Abstract
The emergence of LLMs necessitates a re-evaluation of communicative agency in social systems. Integrating Habermas’s theory of communicative action with speech act theory, this paper investigates whether LLMs can engage in communicative action, produce speech acts, and function as social agents. We advance the novel claim that LLMs are “asymmetrical communicative agents” that satisfy behavioral but not intentional conditions for communicative action. While LLMs cannot yet generate genuine illocutionary acts, they produce significant perlocutionary effects. We adopt a gradualist position, arguing that LLMs lack a full-fledged “cognitive-volitional” complex, making them susceptible to co-option for strategic action even as they intervene in canonical communicative scenarios. This dual-status ontology suggests LLMs are quasi-social agents capable of advancing validity claims and influencing social norms. After anchoring LLMs within a “digital lifeworld,” we analyze their evolving roles across the strategic-communicative spectrum and their implications for the construction of social reality.
1 Introduction
This paper interrogates the status of communicative action in the age of AI dominance. As developed by Jurgen Habermas, the concept of communicative action integrates several strands of social action theories arising within sociology and theories of intentionality and language arising primarily within philosophy (Heath, 2001). In particular, Habermas draws from the sociology of Mead, Parsons, and Weber and the schools of phenomenology, critical theory, and speech act theory in order to distinguish between two chief mechanisms of social action: strategic action and communicative action (Habermas, 2007a). As Heath points out, this distinction loosely mirrors Weber’s distinction between purposive-rationality [zweckrationalitat] and value-rationality [wertrerationalitat] (Heath, 2001). Weber defined purposive-rationality as instrumental, purposive and goal-directed rational action where a social actor weighs the means to their “calculated ends”, and value-rationality as value-oriented rationality where action is constrained by a value system of “ethical, aesthetic, religious, or other forms of behavior, independently of individual ends” (Weber, 1978).
The concept of instrumental reason and the means-ends schema trace at least as far back as Aristotle’s teleological theory of action (Aristotle, 1995). In early modern philosophy, the instrumental conception of rationality shows up in Hobbes’ Leviathan where he states that “for the thoughts, are to desires, as scouts, and spies, to range abroad, and find the way to the things desired” (Hobbes, 1996), and more pejoratively in Kant, where he formalizes a sharp distinction between the autonomy and the heteronomy of the will (Kant, 2002b). Whereas Hobbes subordinates reason to the means of the ends supplied by “the passions” (Hobbes, 1996), Kant relegates instrumental reason to “inclination” subsumable within a deterministic picture of nature and elevates the will that observes universalizable rules to “autonomous” action that escapes the necessity of antecedent causes (Kant, 1996, 2002a). These two ostensibly conflicting conceptions of reason, one as instrumental, and the other as fundamentally cooperative and “socialized”, have vied for dominance in 20th and 21 st century thought, with economics and game theory championing the former, and certain schools of sociology and philosophy, the latter (Parsons, 1949; Habermas, 2007a; Eriksen & Weigård, 1997; Heath, 2001).
Within this context, Habermas’s theory of communicative action represents the culmination of a lineage of ideas that privileges non-instrumental reason as a constraint on the avarice of instrumental reason. Communicative action constitutes the chief social mechanism whereby the social world laden with norms, values, and rules that humans inhabit acquires validity and legitimacy. Habermas defines communicative action as linguistically mediated social action whose concrete aim is to achieve mutual understanding between communicative agents, rather than attain ends (Habermas, 2007a). By contrast, Habermas dubs strategic action the types of social actions, whether linguistically mediated or not, that maintain a success-orientation, encompassing within its purview the longstanding teleological conception of action and rationality (Habermas, 2007a). Habermas carves distinct sociological spheres of operation, though by no means mutually exclusive, for these two overarching mechanisms of social action: whereas communicative action subserves the lifeworld, strategic action subserves the system. By lifeworld, Habermas designates the background of shared understandings and practices that enables individuals to navigate the social environment, whereas by system, the formal structures spanning governmental, administrative, legal, and economic powers that organize society (Habermas, 2007b). Habermas pins the unmooring and disentangling of these once concomitant spheres into separate vectors of development to modern processes of Weberian-bureaucratic rationalization. Whereas the lifeworld operates through mutual understanding, the system evolves through processes of differentiation. This theorizing of the gradual separation of lifeworld and system, presided over in the 20th century by the Weberian bureaucratic model, warrants reexamination in light of the dominance of ICTs and artificial intelligence (AI) in our milieu. This paper rethinks the structural relations between the lifeworld and system in what has been dubbed as the “digital lifeworld” by analyzing the emergence of human-LLM communication as an entirely new paradigm of communication that upends orthodox assumptions about the successful conditions of linguistic communication.
For Habermas, the very possibility of linguistic communication attests to a primordial agreement and consensus orientation undergirding social action that precedes teleological action (Habermas, 2007a). As a precursor to his notion of communicative action, Habermas draws from the symbolic-interactionist sociology of Mead, who takes the “social act” as prior to, and the larger context for, higher-order phenomena like language, consciousness and mind (Mead, 1934). For Mead, the social whole precedes, and consequently conditions and contextualizes, the behavior of individuals within groups (Mead, 1934). Similarly, Habermas recognizes that “instrumental actions are set within cooperative interrelations of group members and presuppose regulated interaction” (Habermas, 2007a), where “only the communicative model of action whereby speakers and hearers, out of context of their preinterpreted lifeworld, refer simultaneously to things in objective, social, and subjective worlds in order to negotiate common definitions of the situation” (Habermas, 2007a). This presupposition that successful communication presupposes subjects with minds and subjectivities finds robust contemporaneous challenge in the emergence and increasing prevalence of human-LLM interaction where, we claim, formal pragmatics diverges from the stringent conditions imposed by both speech act theory and certain interpretations of the theory of communicative action. In this paper we therefore challenge the condition of the “intersubjective recognition of linguistic claims” (Habermas, 2007a), to varying degrees underlying both Habermas and Searle’s conception of speech acts, as the essential condition that enables communicative action and a fortiori, speech acts by advancing, what we call, behavioral conditions for both speech acts and communicative action that paradigmatically define human-LLM interaction.
The impetus of this paper derives from the observation that the “standard conditions” of communication are rapidly shifting with the onset of human-LLM communication, and that the conditions for communicative success, whether illocutionary or otherwise, are shifting underneath our feet. It is further motivated by the exigent concern that, on the one hand, the philosophically state-of-the-art models of social action such as those spawned by Habermas promulgate a norm-bound alternative that regiments and tempers unfettered instrumental action, and on the other, that the contemporary shape of sociotechnical systems with AI as the frontier of the computational paradigm, are hollowing out the processes whereby communicative action occurs and garners social force. Recent literature published after our initial drafting further supports this position, framing LLMs as colonizers of the lifeworld that deterritorialize surviving regions of communicative action by replacing it with instrumental action (Martineau et al., 2026). However, in this paper, we take a more balanced approach centered on the argument that LLMs can straddle both communicative and instrumental action. As such, we begin by first outlining how LLMs fit within an increasingly dominant digital lifeworld; second, we consider their capacity for speech acts and how they fit into linguistically delegated chains of action that produce collective intentionality; third, we outline how LLMs can be and are co-opted as instruments of teleological and strategic action; and fourth, we consider how LLMs fit into the ideal communicative action scenario by positing the novel category of asymmetrical communicative action.
2 AI & the Lifeworld
The recent advent of deep learning and more recently LLMs raise acute questions about the status of the distinction between lifeworld and system, on the one hand, and their relations, on the other. Should we think of AI and LLMs more specifically as belonging to the system or the lifeworld or both in distinct ways? Further, does AI present prospects for abridging this distinction or does it threaten to collapse the lifeworld into the system? The answers to these questions, we will argue, lie with understanding concretely AI’s influence on the shape of strategic and communicative action. Since strategic action lends itself to the system whereas communicative action to the lifeworld, the predominance of one over the other will be indicative of the shape of their distribution. An adequate theory of human-LLM relations should therefore contend with the space that LLMs currently, and increasingly, will occupy in the intersubjective domain of shared social communication and understanding, dubbed the lifeworld. To bring this theorization step to contemporaneous speed, we first situate Habermas’s deployment of the conception in the antecedent phenomenological conceptions of Husserl and Heidegger, and subsequently consider the status of LLMs in the recently theorized notion of the “digital lifeworld”.
Habermas situates his theory of social action as bifurcating between strategic and communicative rationality within the broader concepts of system and lifeworld. Drawing primarily from Husserl, Habermas defines the lifeworld as “the conditions under which the unity of an objective world is constituted for the members of a community … [which] gains objectivity only through counting as one and the same world for a community of speaking and acting subjects” and where “the abstract concept of the world is a necessary condition if communicatively acting subjects are to reach understanding among themselves about what takes place in the world or is to be effected in it” (Habermas, 2007a). He further qualifies that “through this communicative practice [the community] assure themselves… of their common life-relations, of an intersubjectively shared lifeworld… bounded by the totality of interpretations presupposed by the members as background knowledge” (Habermas, 2007a).
Focusing less on its communicative dimension and more on its givenness, Husserl defines the lifeworld [lebenswelt] in his The Crisis of European Sciences and Transcendental Phenomenology, Husserl as:
The ontic meaning [Seinssinn] of the pregiven life-world* is a subjective structure [Gebilde], it is the achievement of experiencing, pre-scientific life. In this life the meaning and the ontic validity [Seinsgeltung] of the world are built up—of that particular world, that is, which is actually valid for the individual experiencer. As for the “objectively true” world, the world of science, it is a structure at a higher level, built on prescientific experiencing and thinking, or rather on its accomplishments of validity [Geltungsleistungen].(Husserl, 1970).
Husserl’s definition of the lifeworld with its emphasis on the predicates of “pregiven-ness” and “pre-scientific” as its essential dimensions, draws significantly from Heidegger’s conception of world expounded in the chapter “The Worldhood of the World” [Die Weltlichkeit der Welt] in Being and Time (Heidegger, 1962). Heidegger distinguishes between four definitions of world: (a) “world as the totality of entities”, (b) “world as a domain of discourse e.g. world of mathematics, world of biology etc.” (c) “world as the concrete set of relations wherein the human being lives” and (d) “world as the ontologico-existential concept that makes possible the other three meanings of world” (Heidegger, 1962), taking the third designation as the subject of his chapter, which spans the ‘public we-world’ and ‘one’s own closest domestic environment’ (Heidegger, 1962). Heidegger’s concept of world must be understood in context of the existential structure of being-in-the-world. Being-in-the-world posits the mutual constitution of agent and environment such that the agent cannot be described independently of its comportment to some environmental context of action (Heidegger, 1962).
This mutual constitution entails that the agent encounters the environment not as an external artifact or a container of entities but foremost as salient contexts of action within totalities of referential networks (Heidegger, 1962). Given Heidegger’s privileging of the action-context, he distinguishes between readiness-to-hand and presence-at-hand as pretheoretical modes of comportment toward entities. Readiness-to-hand signifies a more primordial and direct relationship to objects. As such, Heidegger argues that the agent is “always-already” in some mode of contextual disclosure that delineates the horizon of salient actions (Heidegger, 1962), which Habermas interprets as a pre-defined and culturally laden world that agents encounter as already interpreted (Habermas, 2007a).
Notwithstanding differences of nuance between Husserl’s and Heidegger’s conceptions, Habermas coalesces common insights that include the constitution of the lifeworld as pregiven and as pre-theoretical. Pregiven means that social actors inherit a world of public meanings as the background for their own socialization and norm adoption. Pre-theoretical indicates that immersion within, and engagement with, the lifeworld precedes thematic articulation and theorization. Social agents navigate the public signifiers of the lifeworld as “second-nature”, a given network of public meanings that encapsulate both descriptive components as well as permissible action-contexts.
Recent scholarship has rightly drawn the concept of the lifeworld into dialogue with the contemporary digital milieu and more broadly, information-laden dimensions of social being. Hybrid extensions such as “digital lifeworld” and “onlife” gesture to integrate the ontological expansion of the lifeworld through information-communication systems (ICTs) and digital fora as competing spheres of social action. By “onlife” Floridi designates an emerging milieu where the integration of ICTs in human society blurs the distinction between “real and virtual” and “human, machine, and nature” (Floridi, 2015). The neologism “onlife” naturally suggests a transformation of the lifeworld from predominantly mediated through local in-person interactions to one where the virtual is co-present with, and co-constitutes, real-life interaction. However, where some see seamless integration others perceive disjunctions. Durt (2022) argues that full-fledged integration of AI into the human lifeworld requires bridging the gap between machine intelligence, understood within the frame of data processing and syntactic manipulation, and human intelligence, understood within the frame of a lifeworld laden with semantic cultural meaning and understanding. In drawing this distinction, Durt eschews the dichotomous categorization of AI as either object, namely tool, or subject, namely person. Rather, he argues that AI occupies a liminal space between these categories in virtue of intervening in the lifeworld as an actor that behaviorally simulates aspects of human intelligence but actually falls short of higher-order understanding and semantic grounding (Durt, 2022). Echoing Searle’s distinction between syntax and semantics (Searle, 1990), Dreyfus’s argument that machines lack tacit knowledge (Dreyfus, 1992), and drawing from recent scholarship that denies that LLMs possess natural language understanding (NLU) (Bender and Koller, 2020), Durt disanalogizes human and machine intelligence on grounds that data processing qualitatively differs from paradigmatically human processes of understanding, meaning and communication (Durt, 2022).
We concur with Durt in understanding AI, specifically discursive AI such as LLMs and embodied AI such as robots, as altering the constitution of the lifeworld in ways that have yet to be fully understood and determined as technology advances. However, contra Durt, we argue that his eschewal of the dichotomous designation of AI as either object or subject can be extended to cybernetic systems more broadly, encompassing computation, ICTs, distributed processing, IoT, and symbolic AI. At the same time, by relegating AI to “syntactic manipulation” Durt fails to address the qualitative distinction between Symbolic AI, which is reliant on explicit symbol manipulation, and connectionist AI, which learns patterns in virtue of its distributed and hierarchical network structure (Kumar, 2024; Rani et al., 2023). More problematically, Durt resuscitates the syntax-semantics distinction, in effect reifying “human meaning” as eliding analysis in terms of “data” or “information-processing”. Connectionist AI, and word embedding algorithms in particular, blur the distinction between syntax and semantics because they learn both as contextual patterns nested in multilevel linguistic representations spanning subwords, words, sentences, documents and text corpora (Clark et al., 2019; Ghannay et al., 2016; Jawahar et al., 2019; Peters et al., 2018). While word-embeddings and dense-vector representations of text data differ in important ways from the ontogenesis of linguistic meaning in human languages (Jones et al., 2024), they also structurally approximate the representation of semantic memory in biological neural networks (Elder et al., 2025; Kauf et al., 2024; Kumar et al., 2024). For example, transformer attention heads have been shown to perform linguistically interpretable operations such as resolving the direct object of a verb phrase (Kumar et al., 2024). While these considerations fill a gap in Durt’s analysis with respect to the distributional semantics of word-embeddings, making LLMs much more mutually intelligible with human agents in the semantic sense than Durt acknowledges, they do not signal an overcoming of the fundamental syntax-semantics gap beleaguering artificial agents, especially given that LLMs lack extensionality and real-world reference.
Offering a contrasting view, Noller (2024) argues that “AI and other technology” are not separate from the analog lifeworld, but rather merge together “with regard to the connection of teleological patterns” into an integrated structure he terms the “digital lifeworld”. Like Durt’s recourse to the syntax-semantics distinction, Noller’s thesis of an integrated lifeworld subsuming AI and digital information systems, resuscitates the common view of AI as forming an extension of human goals. He states that “ANNs…can be understood in terms of abstracting patterns by means of machine learning and applying them to new teleological environments” (Noller, 2024). In construing artificial neural networks (ANNs) as extensions of human purpose and teleological patterns of action, he relegates them to “tools” that augment human instrumental action. In our terminology, this view construes ANNs and LLMs as means within structures of human strategic action. However, this conception neglects to address the constitutive dimension of the lifeworld as the “background knowledge” and implicit norms that condition contexts of purposive action. In understanding the lifeworld exclusively with respect to “contexts of purpose”, Noller sidesteps the question of how human-LLM interaction can shape and transform the pretheoretical, tacit givens that undergird social norms.
Whereas Durt emphasizes persisting friction and discontinuities between the human lifeworld and AI as a paradigm of digital information processing, Noller predicates their integration in the subsumption of AI within human contexts of action. However, what both views fail to address in their theorizing of the “digital lifeworld” is the constitutive role of digital technologies in shaping the “pretheoretical” and norm-laden sphere of action, on the one hand, and the explicit and discursive negotiation of norms through communicative action. For example, neither addresses the discursive capabilities of LLMs as a behavioral modality that fundamentally alters human-technology relations. Agreeing with Durt that AI should not be ontologically classified into an either/or schema as “person” or “tool” and with Noller that digital technologies enmesh with the human lifeworld, we fill in two important gaps in their analyses: (a) describe the effects of human-LLM interaction on the pretheoretical, norm-constitutive aspects of the lifeworld and (b) take LLMs as quasi-social actors with sufficient discursive capacities to shape explicit communicative action. Interpreting the lifeworld as the social sphere of action where linguistic meaning plays a constitutive role in informing norm-bound behavior as well as its explicit extension through communicative action, we argue that human-LLM interaction marks an ontological shift in the negotiation of implicit norms and generation of explicit consensus within the lifeworld.
In particular, we argue that LLMs as discursive agents possess new causal powers in shaping the lifeworld. Whereas prior ICTs and digital infrastructures such as social media structurally reconfigure the lifeworld at the level of system, LLMs possess the capacity to alter the lifeworld at the level of communicative action. In other words, recent theorizing of the digital lifeworld analytically blurs the distinction between system and lifeworld and neglect to factor in the role of the discursive prowess of LLMs in affecting interaction at the scale of individual agents and partaking in explicit processes of reasoning and norm justification. Whereas Habermas theorized that the system hollows out the lifeworld (Habermas, 2007b), we argue that sociotechnical systems structurally coalesce system and lifeworld into digital social action, wherein human action is increasingly input to a digital recording system. However, while structurally concurrent, we maintain that system and lifeworld remain independent vectors that carry distinct operations within the digital lifeworld. For example, the digital lifeworld of reddit users is laden with the norms of the subreddits they frequent, the types of interactions those norms foster, and the type of actions the platform affords.
As such, we propose that digital action constitutes the unit of analysis that simultaneously embodies both system and lifeworld vectors. When a user engages in “upvoting”, “downvoting”, “posting” and “commenting”, these constitute digital actions that interact directly with the lifeworld of a digital ecosystem. These same digital actions, however, generate metadata that accrue to the user’s profile and subsequently serve as inputs for the generation of recommendations, personalized ads, and structure their experience of the platform. In other words, two sets of interaction parameters coincide: (a) the action-possibilities defined by platform affordances, and (b) the units of record that serve as inputs into the platform’s algorithmic regime. These two sets of interaction parameters often overlap, but not always. For example, explicit inputs such as clicks, ratings, likes, votes, shares, and follows simultaneously count as user interactions and system activity logs. On the other hand, while more implicit activities such as hovering, session-length, skips, visit frequency, dwell-time, scroll-depth, may not constitute significant user actions, they nonetheless contribute to the user’s activity log and digital footprint. The co-presence of system and lifeworld in digital worlds constitutes a novel dimension of practical reason because every user action is potentially an input to an extractive system that aims to model user behavior not only to predict it but also influence and disseminate its activity pattern. Consequently, even though not all digital actions are social actions, digital actions are increasingly system actions. While this attests to the mediation of digital action through large platforms and software-as-a-service that make participation conditional on identity authentication, growing communities of users actively resist these hegemonic practices through open-source ecosystems, self-sovereign identity and hacker culture (Söderberg & Maxigas, 2022; Webb, 2020). The bifurcation of the digital lifeworld into frictionless platform participation and effortful resistance marks the cultural fault line along which the digital public sphere is contested. The possibility of the coincidence of social and system actions differs fundamentally from action outside of the digital lifeworld where these two domains can become uncoupled.
Given the increasing dominance of digital action as the primary means through which individuals interact with the lifeworld, this analysis naturally extends to LLMs. LLMs embody the system properties we enumerated above by constituting digital recording systems whose user outputs form inputs to further training and modeling. However, LLMs also embody novel properties in virtue of their unprecedented linguistic competence. While publicly accessible LLM chatbots are on the surface currently passive agents (at least in non-agentic scenarios) that require user prompts to initiate interaction, the interaction patterns they enable both emulate and far exceed the possibilities of interpersonal human-human communication. These general language capabilities in conjunction with the computational pooling of vast ranges of human knowledge enable interactions that span three core functions of human linguistic competence: (a) representational (b) interactional and (c) personal. Representational communication encompasses interactions related to knowledge retrieval and transmission, reasoning, belief-checking and updating. Interactional communication encompasses interactions that fulfill the need to socialize as an end in itself without an instrumental goal in mind such as sharing, feeling seen and heard, and alleviating loneliness. Personal communication encompasses interactions such as confiding sensitive personal information and the general sharing of feelings and emotions.
In our theoretical analysis, we advance five distinct theses as novel contributions to the sociotechnical understanding of the digital lifeworld:
-
(a)
The digital lifeworld is marked by the contiguity and functional re-integration of lifeworld and system.
-
(b)
Digital action is a distinct type of social action that leaves a digital imprint on recording systems.
-
(c)
LLMs constitute quasi-social agents that perform functional speech acts, which contribute to and extend human collective intentionality.
-
(d)
LLMs constitute hybrid tool-agents that additionally subserve strategic rationality.
-
(e)
LLMs constitute asymmetrical communicative agents that shape implicit norms and partake in explicit consensus generation in communication with human agents.
Given that linguistic meaning and language use are constitutive of the lifeworld, the role, status, and effects of LLMs within the realm of linguistically-mediated social action needs to be charted. Not only do LLMs alter the presupposed conditions of successful communication, but they introduce a new notion of social action that we dub LLM-action, not reliant on a “cognitive-volitional complex”. Habermas defines the cognitive-volitional complex as the internal state of an agent, encompassing their knowledge and will, which enables strategic action, but is insufficient on its own to explain communicative action, reliant on a shared, intersubjective understanding that transcends individual agents (Habermas, 2007a). Consequently, LLM actions can be thought of as means of human ends or as collaborative efforts that are shaped through continuous human-LLM interaction. The disposition to collaborate, “hard-wired” within LLMs through alignment techniques, does not preclude the fact that human-LLM collaboration emerges from the dialogical pattern of their interaction, which is reciprocal in the sense of being dialogical and not in the strong communicative action sense of implying an overlapping intersubjectivity. Language structures thoughts and conceptual understanding, which form prerequisites for social action and cooperation. For Habermas, language acquisition is tantamount to enculturation into a “preinterpreted domain of what is culturally taken for granted” (Habermas, 2007a). While language acquisition is a largely verbal process, verbal and textual LLMs substitute an artificial interlocutor that inscribes both domains of knowledge and cultural norms. In the rest of the paper, we proceed by explicating the three roles we have designated for LLMs in the lifeworld: (c), (d) and (e).
3 LLMs & Collective Intentionality
In this section, we develop novel implications that LLMs have for collective intentionality. Intentionality is a philosophical term of art that refers to the property of minds to be directed at, or be about, states of affairs (Searle, 1980, 1995). Collective intentionality, on the other hand, refers to a species of intentionality underpinning collective behavior, through which agents participate in coordinated action by virtue of sharing mental states or certain behavioral dispositions (Gilbert, 2014; Searle, 1990). One canonical analysis of intentionality ascribes it the logical structure S(p), where S denotes the intentional state and p its contents (Anscombe, 2000; Searle, 1969). For example, a belief intentional state can have the content that “Paris is the capital of France”. Although intentional states need not have linguistic contents, propositional contents expressed in natural language enable a species of linguistic action termed speech acts that are constitutive of social reality (Searle, 1969). Consequently, in this section, we consider the degree to which LLMs approximate linguistic behavior that constitutes functional equivalents of intentional contents and effects. The propositional content expressed in natural language allows mental representations to be shared and synchronized, raising the question whether LLMs can participate in this synchronization through speech acts. Getting clear on these questions is of paramount importance given the increasing pervasiveness of human-LLM interaction (Elon University, 2025) on the one hand, and the increasing role of LLMs in human-in-the-loop and humans-out-of-the-loop decision-making processes on the other (Mellamphy, 2021). As we will argue, individual human-LLM interaction augments individual intentionality, yet at the same time carries implications for collective intentionality. Further, LLMs and other forms of state-of-the-art AI mark an inflection point in algorithmic decision-making on account of approximating human reasoners. The extent to which LLMs can count as human reasoners will have profound implications for the role that LLMs play in collective action. Despite arguments that LLMs cannot count as full-fledged human reasoners because they do not possess reflective knowledge (Hila, 2025), evidence shows that LLMs can function as reliable transmitters and evaluators of existing knowledge (Havrilla et al., 2024). To count as human reasoners LLMs would require reflective knowledge, namely possess internal access to the bases on which beliefs, or functional equivalents thereof, are justified. Reflective knowledge is an epistemological theory that privileges an internalist standard for knowledge, wherein reliable belief causing processes are necessary but insufficient conditions for knowledge unless those processes are internally evaluated and rationally justified.
3.1 Theories of Collective Intentionality
Before defining the relations between LLMs and human collective intentionality, we begin by providing context for the contemporary debate on collective intentionality. These debates center on the tension between the thesis that (a) collective intentionality is irreducible to an aggregation of individual intentional acts, and (b) that individuals nonetheless possess individual intentionality. This creates a puzzle about how collective intentionality is possible given that: (a) collective intentions cannot be reduced into sums of individual intentions, yet (b) all intentions are instantiated in individual minds.
The coinage of the technical term “collective intentionality” traces to John Searle, who defined it as a collective intention-in-action goal whose means are the individual intentions-in-actions of the parties involved (Searle, 1990). Searle defines an intention-in-action as a type of self-referential causation where an intention’s representational contents cause those conditions to obtain in the world: e.g. I feel thirsty, so I get a drink (Searle, 1995). According to Searle, “We-intentions cannot be analyzed into sets of I-intentions, even I-intentions supplemented with beliefs, including mutual beliefs, about the intentions of other members of a group” (Searle, 1990). Searle solves the puzzle of the insufficiency of collections of I-intentions constituting We-intentions by arguing that We-intentions are primitive or irreducible intentional states that individuals can hold (Searle, 1990). Specifically, Searle argues that individual intentions-in-action of the participants involved posses the form: Collective (goal) B by means of singular (actions) A, where the individual actions serve as means to, or constituents of, a collective goal (Searle, 1990). In other words, while the intentions are individual, participants represent them through irreducibly collective contents, where the means do not constitute separate actions, but constituents of a collective action. In this structure, the individual action (A) is not an external cause of the collective goal (B), but is constitutively nested within it, functioning as the specific means by which the agent participates in the collective whole. Searle’s analysis is indebted to Sellars (1964), whose earlier theory of collective intentionality argued that we-intentions are individual, but their contents shared amongst participants.
Gilbert (2014) offers a competing account to Searle’s theory of collective intentionality based on joint commitment. Gilbert defines joint commitment as follows:
“Persons X and Y collectively intend to perform action A if and only if they are jointly committed to intend as a body to do A” (Gilbert, 2014).
Gilbert takes commitments to entail “sufficient reasons” to act in accordance with what has been committed (Gilbert, 2014). As such, commitments are “commitments of the will” whereby acting contrary to a commitment, unless it is rescinded, entails a contradiction of the will. Because commitments imply sufficient reasons to act, they bind persons to carry out their intentions. In the above definition, Gilbert qualifies “intend as a body” to mean that a “joint commitment” is irreducibly singular in the sense that it cannot be decomposed into an aggregation of individual commitments. How does one go from individual commitments and intentions to joint commitments and intentions? For the latter to be possible, Gilbert defaults to a communicative condition whereby “matching states of the will” must be “mutually expressed” or “become common knowledge between the parties” (Gilbert, 2014). Because joint commitments share the logical structure of individual commitments, they provide sufficient reasons for each party to satisfy them. However, Gilbert’s account relies on taking her “communicative condition” for joint commitment to fall short of “agreement”, making the latter a special case of joint commitment. For this to hold, we would need to sharply distinguish between “tacit agreement” and “explicit agreement”, and good reasons can be provided for taking any communicative condition of mutual understanding as existing on a spectrum of “explicit agreement”.
By positing joint commitment as a precondition for “collective action” (Gilbert, 2014), Gilbert solves the ostensible puzzle of collective intentionality by retaining its irreducibility to individual intentionality while explaining how singular agents can partake in collective action as a body (Gilbert, 2014). Specifically, when commitments to intend to jointly act turn into actions, the result is not a sum of individual wills but rather a “collective agent” or a “supra-individual agent” (Gilbert, 2014). At the same time, Gilbert views this “collective agent” as comprising members with individual plans and rational capacities to rationally evaluate their joint commitments and participation in collective action. Gilbert’s positing of “collective agents” contrasts with the views of Sellars and Searle, who maintain that intentions are ontologically individual, but the contents are irreducibly plural (Searle, 1990).
Parallel to Searle and Gilbert’s ontological accounts of collective intentionality, Tomasello adopts a developmental-evolutionary approach, tracing the origins of full-fledged intentionality as imbued with a belief-desire psychology in early forms of intentionality exhibited as early as 6 months to 1 year period of child development (Tomasello, 2014; Tomasello and Rakoczy, 2003). Tomasello hypothesizes that behavioral evidence of joint-attentional abilities points to a primitive type of social-cognitive apprehension of others as “intentional agents” in early development that is unique to humans (Tomasello and Rakoczy, 2003). Tomasello et al. (2003) further advance the claim that human understanding of other persons emerging in infants forms a precondition for linguistic and other types of social learning that encompass human culture including social norms and collective beliefs, which they equate with collective intentionality. This “species-unique form of social cognition” enables children, for example, to establish self-other equivalence (Tomasello and Rakoczy, 2003). Tomasello’s empirical investigations echo Searle’s hypothesis that collective intentionality presupposes a “preintentional sense of ‘the other’ as an actual or potential agent like oneself in cooperative activities” (Searle, 1990). More precisely, Tomasello supplies evolutionary-adaptive grounds for the innate human recognition of others as intentional agents as the cognitive scaffolding for the socialized development of the belief-desire psychology upon which Searle and Gilbert base their theory of intentionality.
These three leading theories of collective intentionality as embodied in Searle, Gilbert and Tomasello jointly supply both ontogenetic and ontological accounts of collective intentionality. However, where Searle and Gilbert differ in their solution to the ontological puzzle, Tomasello provides an empirical basis for understanding the ontogenesis of collective acceptance and action rooted in the adaptive trait of infants to infer and distinguish intentions in others, both communicative and otherwise, from mere action outcomes. This primitive basis for recognizing communicative intentions provides a neat link to Habermas’s theory of communicative action, which vehemently affirms mutual understanding as the basis for both sociality and the emergence of language, the latter of which cannot arise from instrumental action alone (Habermas, 2007a; Heath, 2001). Tomasello shows that from early development, individual infant intentionality is inextricably bound with shared intentionality with others, whence socialization of norms and linguistic competence arise. Vis a vis communicative competence, Tomasello argues that social acts such as pointing or making a sound to draw shared attention are intersubjective in that the child learns the symbolic form of the communicative act that it then replicates for the same purpose (Tomasello and Rakoczy, 2003). In the case of norms, deliberate misnaming games provide evidence that children understand the normative force behind naming conventions (Tomasello and Rakoczy, 2003). Consequently, Habermas’s theory of communicative action can be viewed as another type of theoretical solution to the ontological puzzle of collective intentionality as well as a theory of social action that is both consistent with, and explicitly supported by, Tomasello’s empirical findings. For Habermas, communicative intentions are primordially oriented towards mutual understanding and consensus rather than being the mere effects of success-oriented (strategic) action. By baking-in a cooperative orientation within social action, Habermas blurs the line between I-intentions and We-intentions, rooting both in the communicative pragmatics of cooperation.
3.2 How LLMs Integrate into Collective Intentionality
These analyses motivate a two-prong account of LLM collective intentionality rooted in (a) the largely innate and invariant intersubjective sources of human collective intentionality and (b) the variable forms it takes within social norms through mechanisms of social agreement and convention generation. The empirical insights outlined provide a basis for explaining why human agents could be predisposed to extend communicative parity to LLM agents given that the communicative imperative is both bound to intersubjective intentions and constitutes the socialized means of learning and validating social norms. More broadly, they provide a basis for understanding human agent dispositions toward competent linguistic agents such as LLMs and the role the latter can play on the social convention-based dimensions of collective intentionality. Our analysis reveals that mutual attribution of intentions undergirds collective intentionality, but given that such intentions are inferred from behavioral cues, it does not preclude the attribution of such intentions to artificial agents. In the case of LLMs, the behavioral paradigm of relevance concerns linguistic competence and the way that linguistic communication figures in processes of consensus generation, the generation of social conventions, and the fulfillment of personal and social needs. Given that linguistic performance and agreement constitute the mechanism for erecting cultural norms, conventions, and rules for both Searle and Habermas, the emergence of human-LLM communication and subsequent LLM integration into social and organizational processes underscores a vector of cooperation and collective behavior that transcends mere human collectivity. We argue that the wedge that absorbs LLM behavior into collective intentionality resides with speech acts: linguistic utterances that perform an action rather than merely describe the world (Habermas, 2007a; Searle, 1969). The possibility of LLM speech acts, we argue, elevates LLMs into agents of collective behavior rather than mere means to ends, namely, as tools. However, we distinguish between intentional speech acts, which presuppose a belief-desire psychology, from LLM speech acts, which, with appropriate authorization, can carry some of the causal powers and behavioral effects of intentional speech acts. Given the foregoing considerations, we posit that LLMs can play a dual role within collective intentionality: (a) through participation in normative communicative action, which we address in the last section and (b) through delegation of social roles that partake in the authorizations of norms, rules, and social facts.
Having explicated the central role of “shared and joint intention” within human collective intentionality, the question we examine next is whether LLMs can play a constitutive role in social reality by interacting with human intentions and joint collective behavior, or whether they play merely a strategic role as mere “means” that subserve human purposes. This question is not merely an empirical question about what LLMs currently do, but rather a theoretical question about what social powers LLMs are capable of in virtue of their linguistic capacities. John Searle’s theory of institutional facts provides a useful starting point for considering this question. Searle distinguishes between “brute” facts and language-dependent institutional facts (Searle, 1995). To illustrate, that money is a unit of exchange is an institutional fact, whereas the distance between the earth and the moon is a brute fact. For Searle, the generation of institutional facts depends on the ability of agents with intrinsic intentionality or minds to assign functions to naturally occurring and made objects (Searle, 1995). Because functions are impositions by intentional agents, they constitute observer-relative descriptions over and above an entity’s intrinsic properties (Searle, 1995). Searle distinguishes between agentive functions that are assigned to objects made to serve a particular purpose such as a chair and non-agentive that are assigned to naturally occurring phenomena such as a heart. Only a subspecies of agentive functions that Searle dubs status functions generate institutional facts by virtue of defining constitutive rules that enable entities to perform a role in a collectively defined context. The closest definition that Searle offers for status functions is that they involve the collective acceptance that some phenomenon represents or stands for something else (Searle, 1995). In this picture, language itself constitutes a nested series of status function impositions that subsequently become the constitutive basis for higher-order assignments such as social roles and procedures (Searle, 1995).
Specifically, Searle pins the mechanism for generating institutional facts to a class of speech acts he terms “declarations” (Searle, 1969, 1995). For instance, the declaration “the meeting is adjourned” makes it the case that the meeting is adjourned. However, not all declarative utterances generate institutional facts. For a declarative speech act to produce an institutional fact it must meet the antecedent condition of the imposition of status function to some entity or person (Searle, 1995). The imposition of status function requires the collective acceptance by intentional agents of the assignment of a role through the constitutive rule of the form “X counts as Y in C”, where x denotes the entity, individual, or event, Y the social function and C the surrounding context (Searle, 1995). An example of such a rule could be: “A person who holds a Juris Doctor (JD) degree and has passed the bar examination counts as a person authorized to give legal advice in that State’s legal system when admitted to the Bar and maintains good standing.” A common criticism of Searle’s theory concerns the explanatory gap between institutional and sub-institutional constitutive rules. For example, the collective acceptance of some object as a unit of exchange can occur naturally without a formal body designating it as such. This means that the assignment of function can be informally binding through tacit assent, or it can be contested altogether rendering its functional status ambiguous. This suggests that despite gestures towards a robust theory of collective intentionality by Searle, Gilbert (2014) and Tomasello (2003), the requisite processes that produce collective acceptance remain largely opaque (Tuomela, 2003).
Because Searle’s theory of collective intentionality and institutional facts relies on intentional agents, it raises the question of the range of possible status functions that can be assigned to LLMs. To answer this question, we draw a distinction between agents that can confer status functions and entities to which status functions are assigned. Searle’s theory holds that only intentional agents through collective we-intentions can confer status functions. On the other hand, status functions can be attributed to both intentional and non-intentional entities, spanning persons, objects and events (Searle, 1995). When status functions are conferred on other intentional agents, namely persons, they vest authority in such persons to generate institutional facts in their designated role. Persons vested with such authority possess deontic powers to authorize social facts but also inherit obligations they must fulfill. Because we claim that LLMs constitute boundary objects between “tools” and “social agents”, it opens the possibility that LLMs can be assigned both functions of “use” and functions of “social authority”. Functions of use encompass tasks such as customer service, legal document processing, software development, medical research summarization, educational training, financial analysis, to name a meager few. Because agentive function assignments are observer-relative, the range of functions assignable to LLMs cannot be exhausted.
On the other hand, the assignment of roles of social authority traditionally reserved for agents who possess language, self-consciousness, and intentionality poses complex questions about whether the attribution of personhood is a unified category or whether it can be decomposed into functional components. Drawing from prototype category theory, we argue that personhood constitutes a graded category that admits prototypical and non-prototypical members (Dennett, 1988; Lakoff, 1987). In particular, we analyze personhood into three interrelated features: (a) behavioral, (b) intentional and (c) volitional. Behavioral features concern the behavioral distinguishability of AI agents from human agents. Intentional features concern the attribution of beliefs, desires and intentions. Volitional features concern the ability of AI agents to realize their intentions through actions. Of these three components, LLMs sufficiently meet the behavioral condition for natural language use and approximate the intentional attribution of beliefs as argued by Herrmann and Levinstein (2025), even though these are primitive and not full-fledged beliefs attributed to human agents. Given these distinctions, we argue that independent of function assignment, LLMs can play the role of social agent due to their intrinsic abilities for natural language use. We define a minimal social agent as a natural or artificial agent that can engage in competent linguistic interaction. As minimal social agents, LLMs consequently exhibit low-grade levels of personhood due to their discursive capacities and the ability to inhabit coherent personas (Abdulhai et al., 2025; Anthis et al., 2025; Park et al., 2023).
These claims are buttressed by recent evidence suggesting that LLMs exhibit emergent behaviors over and above mere next token prediction and generation (Brown et al., 2020; Lindsey, 2025; Wei et al., 2022; Dutta et al, 2024). Sufficiently large models exceeding 1 billion parameters demonstrate the ability for in-context few-shot learning (Brown et al., 2020). In-context learning constitutes a major breakthrough because it enables LLMs to adapt to context-specific tasks without recomputing the trained parameters (Brown et al., 2020). This general capacity functionally resembles the working memory of cognitive agents that mediates between long-term memory and real-time situational demands (Russin et al., 2025). It differs from working memory in that LLM in-context learning cannot encode new patterns in the model gradient (Brown et al., 2020; Radford et al., 2019). Another emergent feature of in-context learning consists of chain-of-thought prompting (Wei et al., 2022). Wei et al (2022) demonstrated that models with at least 10 billion parameters show significant performance improvements when tasks are broken down into discrete logical steps. These emergent powers have stimulated experiments into whether LLMs have functional introspective access to their own internal states (Lindsey, 2025). Recent evidence shows that sometimes LLMs can reliably report changes to their internal gradient from “concept injection” techniques (Lindsey, 2025). Functional introspective access does not refer to human conscious introspection but, as Lindsey defines it, to the ability of an LLM to accurately detect and report internal, intermediate gradient states (Lindsey, 2025). Correspondingly, the question of whether LLMs possess the capacity for intentional deception is an ongoing area of research (Hagendorff, 2024; Wu et al., 2026).
Together, the emerging paradigms of in-context learning, chain-of-thought reasoning, and introspective AI indicate that LLMs may fit within a spectrum of the cognitive-volitional complex, which Habermas posits as an antecedent condition for strategic action (Habermas, 2007a). While these results do not assign LLMs volitional capacity, they function as indicators of non-anthropomorphic cognitive capacity. Our analysis here differs from that of Martineau et al. (2026), who resuscitate the “stochastic parrot” description (for a nuanced overview of this debate, see Mitchell & Krakauer, 2023). Further, we posit LLMs as social agents on the limited basis of their ability to simulate interpersonal communication with human agents to a high degree of context sensitivity and the ongoing evidence that human agents both engage with them in social interaction and sometimes treat them as sui generis social agents (AbuMusab, 2024; Klein 2025). These emergent behaviors make LLMs strong candidates for social function assignment. Provided that they meet reliability benchmarks, the assignment of social roles would enable LLMs to authorize, adjudicate, ratify, and decide institutional processes. If LLMs can be delegated into social roles where they can perform deontic duties, declarative speech acts and status assignments, they would become causal agents embedded within the explicit structures of collective intentionality.
3.3 LLMs As Performers of Behavioral Speech Acts
Given the foregoing analysis, we now proceed to address the question as to whether LLMs require a cognitive-volitional complex to perform genuine speech acts. Here we offer a novel argument that LLMs perform what we call behavioral speech acts. The category of behavioral speech acts aims to amend Searle’s strict requirement that speech acts necessarily derive from agents that possess original intentionality (Searle, 1980). Because an LLM does not have a genuine intention to produce an effect on a hearer, nor posses a theory of mind about other agents in the strict sense (though it possesses knowledge about human psychology), it cannot perform intentional illocutionary speech acts, namely speech acts that require agents with intentional states. On the other hand, LLMs can perform utterances that are behaviorally indistinguishable from genuine illocutionary acts such as asserting, declaring, expressing, directing and committing. These nonintentional speech acts, which we call behavioral, can produce genuine perlocutionary effects on human hearers. Searle defines the locutionary act as the production of a meaningful utterance, the illocutionary act as the specific force or action performed in speaking (such as promising or ordering), and the perlocutionary act as the actual effect or consequence produced on the listener (such as persuading or alarming) (Searle, 1969). As such, following Searle’s distinction between locutionary, illocutionary, and perlocutionary speech acts (Searle, 1969), we argue that LLMs routinely perform locutionary speech acts that carry perlocutionary effects. While LLMs do not possess “genuine” intentions, they model phrases that inscribe intentions from their training data and generate novel propositions that inscribe “speaker meaning”. For example, an LLM can “apologize”, express “distaste”, ask their interlocutor to “do x”. Because such speech-acts are not tethered to a “cognitive-volitional complex”, we offer a behavioral criterion for LLM speech acts and later for Habermasian communicative action. On the other hand, we acknowledge that LLMs currently fail Searle’s criterion for illocutionary acts (Searle, 1969). Searle (1969) offers the following conditions that agents must meet to perform illocutionary speech acts:
-
(a)
Propositional contents: the utterance must meet the requisite propositional content requirements.
-
(b)
Preparatory conditions: the communication context between speaker and hearer meets the necessary background facts and institutional criteria for the successful performance of the act.
-
(c)
Sincerity conditions: the agent must possess the genuine psychological state expressed by the speech act.
-
(d)
Essential condition: defines the performative role the utterance counts as.
Condition (a) ensures that the speech act meets syntactic well-formedness and semantic content. Condition (b) ensures that the speech occurs in the sanctioned institutional context to carry authoritative force. Condition (c) ensures that the agent intentions align with the propositional content of the speech act. And finally (d) defines the scope of the social state of affairs the speech act makes the case such as an institutional obligation or a declaration that engenders a social fact. It can be plausibly argued that LLMs can meet all the criteria except for (c), which requires that the agent possess psychological intentions, whereas (d) is exclusively a matter of convention that can be extended to LLMs. Habermas contracts condition (b) and (d) into a general rightness condition that specifies whether the utterance is normatively justified in the social world. Consequently, Habermas posits three criteria that function simultaneously as validity claims: (a) a truth or correspondence condition, (b) a rightness condition, and (c) a sincerity condition, whereas a fourth, (d) comprehensibility, functions as an act-invariant precondition of intelligibility (Habermas, 1979, 2007a). Further, unlike Searle, Habermas distinguishes between illocutionary intent oriented toward mutual understanding and illocutionary intent instrumentalized to produce perlocutionary effects. Whereas for Searle the illocutionary intent is successful when its perlocutionary effect on the hearer produces a recognition of its intention, for Habermas the success of the illocutionary intent in communicative action is contingent on the recognition of the validity of the speaker’s content by the hearer. Therefore, Habermas pins communicative action to the illocutionary intent of reaching mutual understanding with the interlocutor, and strategic action to the intention of producing perlocutionary effects on the interlocutor regardless of mutual understanding.
Because LLMs plausibly meet all the criteria except the sincerity condition (c) under both Searle’s and Habermas’s necessary and sufficient conditions for illocutionary speech acts, it follows that LLMs cannot perform genuine illocutionary acts. However, even though LLMs lack intentions, they can in principle satisfy all the behavioral conditions for speech acts. It worth noting here that recently Martineau et al. (2026) have posited that LLMs lack understanding, explainability, and capacity for evaluative judgements. However, these claims are theoretically valenced and rely on cognitivist interpretations of the relevant terms. While we grant that LLMs lack understanding in the sense of conscious thought, whether they possess natural language understanding (NLU) is a deeply contested claim (Mitchell & Krakauer, 2023). The claim that they lack explainability is also a claim about their internal constitution, which equally extends to human brains. For our analysis, what matters is that LLMs can produce cogent linguistic explanations for their claims and serve as generally reliable transmitters of existing bodies of knowledge. Therefore, our argument for behavioral speech acts is not contingent on the anthropomorphic parity of the cognitive mechanisms underlying LLMs but their observed conversational competence and increasing alethic reliability, which is scaffolded through kludges like retrieval augmented generation (RAG). As such, we argue that LLMs can produce quasi- or behavioral speech acts that carry perlocutionary effects on human subjects. We argue that the degree of perlocutionary force depends, in part, on the imputation of “intentions” to LLMs by human interlocutors. However, we further argue that LLM perlocutionary effects on human hearers are also largely independent of the “theory of mind” users ascribe to LLMs and proceed from (a) the degree of alethic trust human agents ascribe to LLMs and (b) because they fit within familiar discursive patterns that human agents embody automatically within norm-bound behavior. The degree of alethic trust refers to the likelihood of truthfulness that human agents ascribe to LLM assertions, whereas discursive patterns refer to a component of trustworthiness that arises from the LLM ability to competently mimic reciprocal discourse, namely their locutionary correctness.
In line with these considerations, we recapitulate our position: LLMs effectuate an uncoupling or severance of illocutionary intent and perlocutionary effect whose relationship is presupposed both by Searle’s speech act theory and Habermas’s theory of communicative action. This uncoupling persuades us to posit the novel category of behavioral speech acts: speech acts performed by artificial conversational agents that meet all the requisite criteria for successful speech acts except for the sincerity condition in Searle’s theory and the cognitive-volitional complex condition in Habermas’s theory, where the former presupposes the latter. We class these conditions as the precondition of intentionality. Because LLMs lack full-fledged intentionality in the sense of possessing (a) psychological states to intend to carry actions and (b) a volitional complex that carries out intended actions, we posit the categories of non-intentional illocutionary acts and non-intentional perlocutionary effects. Because LLMs satisfy the formal conditions for truth and rightness, they can be instrumentalized for individual and organizational strategic action on the one hand, while participating in unidirectional communicative action with human subjects on the other. Our argument that LLMs reproduce the behavioral effects of speech acts in absence of the psychological motivations to either reach agreement or produce perlocutionary effects, makes them amenable to both types of action in different ways. In succeeding sections, we explore both of these possibilities starting first with the effects of LLMs on strategic action and subsequently considering their effects on communicative action.
These considerations further persuade us to advance the novel claim that LLMs can override or circumvent Searle’s transitivity condition for delegating social authority. The transitivity condition, as we interpret it, entails that intentional agents can confer status function only on other intentional agents to authorize social facts. We deny the transitivity condition on functional grounds: intentional agents in principle can assign social functions to LLMs on the basis of their ability to carry out behavioral speech acts that meet requisite alethic benchmarks, namely benchmarks of information-processing reliability. It is important to clarify here the narrowness of the claim: while LLMs can be delegated the authority to inhabit a social role in some circumscribed context, constitutive authority for status function assignment remains vested with human agents with genuine intentionality. As such, LLMs cannot delegate social roles or declare constitutive rules, but only exercise delegated authority in the assigned context. This ensures that the legitimacy of social authority derives from We or collective intentionality. These considerations suggest that LLMs can factor causally within collective intentionality by performing socially delegated roles that produce institutional facts. Figure 1 provides a schematic representation of our proposed revision to Searle’s speech act taxonomy.
Given the potential for delegated action, and in line with our dual-aspect ontology, we now turn to an analysis of LLMs as instruments of strategic action, that is, as means to existing strategic ends.
4 LLMs & Strategic Action
In having argued for, on the one hand, an update and reconsideration of the notion of lifeworld in light of digital environments increasingly laden with artificially intelligent agents, and the potential for LLMs to communicatively contribute to collective intentionality, we now proceed to develop a theory of AI agents as instruments of strategic rationality rather than communicative rationality. We shall address LLM contributions and challenges to communicative rationality in the last section. The previous section established the communicative conditions through which LLMs fit into collective intentionality, namely the performance of functional speech acts absent sincerity conditions. This functional approximation of speech acts forms the condition of possibility for LLM strategic and communicative action. In this section, we shift tone to outline the ways in which LLMs fit into system and action rationality. Habermas argued that the functional differentiation of the system chiefly through exchange media such as currency and organs of administrative power become pathological through what he called reification (Habermas, 2007b). The process of reification entails that the rationality of the societal subsystems such as the economy and public institutions embroil the action rationality of individuals into patterns of action that entirely elide communicative action (Habermas, 2007b). Where the differentiation of the system into specialized spheres follows a rational impetus toward efficient coordination of social ends, reification congeals the rationality of steering media into self-reproducing relations at the expense of the communicative surplus of the lifeworld, effectuating its colonization. In consequence, action rationality, namely the teleological orientation toward individual ends, becomes the chief means through which individuals coordinate with others and encounter the social world, bypassing the consensus-building processes that endow lifeworld norms with intersubjective validity (Habermas, 2007b). In this analysis, action rationality comes to subserve system rationality with the broad effect of emaciating the communicative relations that reproduce the symbolic order of the lifeworld.
In his later work, Habermas integrated the communicative throughput of the internet and the emergence of social media platforms into this analysis by arguing that the new communicative paradigm promulgated by networked platforms dissolves the gatekeeping boundaries of the old mass media reliant on specialists filtering discourse into the public sphere and replaced them with a regime that in principle extends the authorial function to all participants (Habermas, 2023). However, while the decentralization of communicative flows produces a centrifugal expansion of the public sphere, it also condenses flows into self-enclosed communication circuits, undermining the function of the public sphere as an engine of consensus generation (Habermas, 2023). Whereas Habermas saw platforms as destabilizing the social and psychological preconditions for communicative action, Couldry and Mejias (2019) extended Habermas’s colonization thesis by theorizing digital platforms as “technological means that produce a new type of ‘social’ capital’ in the form of personal data extraction. According to Couldry & Mejias (2019) data colonialism entails capitalism’s appropriation of the ‘social’, namely the performance of everyday life, into an extractible resource. These analyses provide a background frame for understanding how human-LLM interaction constitutes an extension of the data colonization project (Martineau et al., 2026), on the one hand, while further reifying the action rationality of individuals toward system ends. As we will argue, LLMs perform a dual extractive function through training data and user interaction, while simultaneously instrumentalizing users into personal utility maximizers to the detriment of authorial and cultural production norms.
Before analyzing the relationship between LLMs and strategic action, it is important to situate LLMs within the cultural ethos that entrenches strategic action in concrete systems such as state administration and corporate structures. As Habermas argues, strategic action as an action-orientation cannot be extricated from historical and sociocultural processes that institutionalized purposive-rational action within state administration and capitalist enterprise (Habermas, 2007b). This framing persuades an analysis of LLMs, on the one hand, as a manifestation or effect of this institutionalization, and on the other, as a means to its augmentation by expanding strategic rationality’s causal powers. Habermas points out that the institutionalization of strategic action obliges individuals in organizational roles to adopt this mode of action (Habermas, 2007b). This downstream effect from institutionalization to normative adoption applies directly to LLMs in a dual sense: individuals are obliged to adopt LLMs to extend both their professional or personal strategic goals and the strategic goals of the organizations in which they are employed. Consequently, our analysis proceeds as follows: first, we consider the requisite conditions that would theoretically qualify LLMs to be strategic actors of their own accord and the degree to which they meet these criteria currently, and second, we advance a structural analysis that frames LLMs as instruments of strategic action in the individual and organizational sense.
In considering the role and effects of LLMs within strategic action, some preliminary definitions are in order. Whether as individual or multi-agent systems, LLMs do not yet countervail incentive structures and teleological ends baked within existing systems of action, but fit ontologically within such schemes of strategic action. We argue that for LLMs to count as strategic actors that realize teleological ends, they must possess a cognitive-volitional complex that forms “beliefs about existing states of affairs through a medium of perception” and “develop[s] intentions with the aim of bringing desired states of affairs into practice” (Habermas, 2007a). Currently, the “cognitive-volitional” complex criterion precludes LLMs from counting as strategic actors because, as we have hitherto argued, they lack both endogenous intentions and means of realizing those endogenous intentions. However, arguments that LLMs approximate the belief-condition in the cognitive-volitional complex have been advanced, even though current philosophical and scientific consensus on whether LLMs possess beliefs is divided (Banerjee et al., 2025; Goldstein, 2024; Slocum et al., 2025). Herrmann and Levinstein (2025) propose four criteria that jointly evaluate whether LLMs possess beliefs: (a) accuracy, (b) coherence, (c) uniformity and (d) use. The first three constraints concern what we call “alethic” conditions, namely conditions that impose veridicality or truth-conditions with respect to states of affairs. Accuracy demands that LLMs possess “true beliefs”, coherence that these beliefs possess logical consistency, and uniformity that truth-representations obey consistency across domains such that beliefs are generalizable across them (Herrmann and Levinstein, 2025). Herrmann and Levinstein’s requirements for consistency and uniformity admit both categorical belief, such as entailment relations, and degrees of belief, namely conformity with a probabilistic calculus (Herrmann and Levinstein, 2025). On the other hand, they define use as true-beliefs playing the belief-role in the LLM’s outputs. This particular criterion marks an ostensible shift from merely alethic conditions to an instrumental condition. The use criterion, in particular, is beset by the “black-box” problem that generally afflicts ANNs, namely the unexplainability of the output given the input (Erasmus et al., 2021). Considered together, however, Herrmann and Levinstein’s criteria only evaluate whether LLMs possess beliefs and whether beliefs causally influence the output. Were LLMs to satisfy all the belief-criteria outlined by Herrmann and Levinstein, they would still fail the volitional component that requires the capacity to develop intentions that bring about desired states of affairs in the world.
Even though LLMs cannot yet generate or act on endogenous intentions, they can carry out the intentions of their designers, trainers, and users. The degree to which LLMs implement and extend the intentions and goals of non-LLM actors determines the degree to which LLMs function as instruments of strategic rationality and action. The absence of a full fledged “cognitive-volitional” complex consigns LLMs to a liminal, boundary object role within the lifeworld. We argue that in producing perlocutionary effects, LLMs constitute a quasi-social actor that influences human actions, beliefs, intentions, and their action-belief complex. On the other hand, because LLMs do not yet possess genuine intentions to effectuate endogenous ends, they are liable to fitting within, and augmenting existing structures of strategic action. Consequently, our analysis proceeds as follows: first, we outline individual and organizational domains of LLM strategic action as the two spheres of action to which LLMs serve as inputs, and subsequently, we analyze the LLM structural components, consisting of training data, pre-training, alignment, and modeling capacities, as both consequences and causes of institutionalized strategic action.
4.1 Human-GenAI Strategic Action
LLMs function as instruments of individual strategic action when they fit into and extend the goals and plans of human strategic actors. The moniker “AI assistant” attests to an increasing cultural understanding of LLM chatbots as augmenting individual goals whether these are confined to productivity or extend beyond professional settings. It is, however, important to qualify that the the evidence on the separation between strategic and communicative uses of LLM chatbots is inconclusive (Kapania et al., 2025; Terzimehić et al., 2025). While data suggests that approximately half of US adults have used LLMs (Elon University, 2025), the degree to which such uses blend functional and social interaction types remains not well understood, not least because general use of LLMs requires dialogical interaction that can seamlessly blend the two types of uses. Whether users view chatbots as teleological extensions likely in part depends on the internal model that users have of the technology, with recent evidence suggesting that users with low AI literacy being more likely to ascribe AI models human-like attributes such as sentience, emotions and agency (Cheng et al., 2025; Huang and Ball, 2024; Tully et al., 2025). At the same time, evidence further suggests that users’ mental models evolve through repeated interactions from exploratory stages to more exploitatory stages, where users shift from an exploration mode that resembles social interaction to an exploitation mode where they use LLMs for personal functions (Holstein et al., 2025; Holstein and Satzger, 2025). In other words, the user’s mental model of the ontological status of the LLM chatbot evolves through ongoing usage. Because chatbots are conversational agents, the line between strategic uses and communicative uses where users seek to come to an explicit normative or alethic understanding of some phenomenon blurs and depends on a host of idiographic factors about the human users such as their personality, profession, social well being and degree of AI literacy. Within this spectrum, strategic uses include using chatbots as personal therapists, schedulers, coding co-pilots, financial advisors, automated researchers, and personalized ghostwriters, to name a few, where the interaction is primarily oriented toward instrumental utility and the optimization of individual, teleological goals. Strategic instrumentalization of LLMs by individual human agents engenders a hybrid form of human-LLM/GenAI strategic action that democratizes professional abilities between human agents, in effect exacerbating the status-game for positional goods. This tendency to maximally exploit LLMs for positional goods paradoxically yields Pareto-suboptimal social outcomes because it intensifies the need for skill differentiation.
4.2 Organizational-GenAI Strategic Action
Strategic extensions of GenAI and LLMs within organizational scales instrumentalizes systemic orchestration agents or components that subserve organizational strategic ends such as maximizing profit and administrative power. As instruments of organizational strategic action, LLMs function as production inputs through integration within orchestration systems of agentic workflows (L Chen et al., 2025; Wu et al., 2024; Zaharia et al., 2024). Organizational workflows loop LLMs within an orchestration process that comprises a data layer, an agentic layer that organizes workflows through the coordination of multiple LLM agents, and a governance layer that ensures cost-filtering, legal compliance and ethical traceability (Dong et al., 2024; Rebedea et al., 2023; Wu et al., 2024). These orchestration frameworks incorporate guardrails such that the aggregation of components and information flows autonomously perform complex tasks that substitute human roles. The transition toward compound AI systems, comprising models and databases, marks a departure from monolithic model usage, situating LLMs as modular components within a broader architecture where systemic performance is driven by the orchestration of diverse tools and retrievers (Zaharia et al., 2024). Through the deployment of multi-agent frameworks, organizations can facilitate autonomous, role-based coordination that effectively automates the ‘strategic action’ and administrative labor previously dependent on human intersubjective communication (Wu et al., 2024). By treating these orchestrated systems as fundamental production inputs, organizations enhance their internal predictive power, thereby optimizing decision-making processes towards organizational bottom-lines (Agrawal et al., 2022). These orchestrated systems not only maximize efficiency but reduce the likelihood of bottom-up communicative action that emerges from human employees, which cause conflicts of interest, and require negotiation and ongoing amendment of organizational workflows to balance bottom-lines with employee well-being and interests.
4.3 Structural Components of LLM Strategic Action
While human-GenAI and organizational-GenAI strategic action describe the effects of GenAI and LLMs on strategic action, it does not provide an analysis of the strategic inputs that LLMs embody. Consequently, the next step of our analysis consists of decomposing the structural components that inscribe corporate and wider cultural forms of strategic action within LLMs. Below we enumerate six distinct ways in which Generative AI functions as an extension of strategic action: (a) Training data and data-access asymmetries, (b) open content production, (c) training and fine-tuning, (d) AI alignment, (e) latent task representation, and (f) conversational agents with the potential for pernicious perlocutionary effects.
4.3.1 Training Data & Data-Access Asymmetries
Given that LLMs do not have independent access to ground-truth, their epistemic horizons are delimited by the training data. Therefore, the degree to which LLMs generate cogent and factual information depends on the cogency and factuality of the training data. Three variables determine the integrity of this pre-computational foundation: the provenance and factual density of the source material, the distributional representativeness of the corpus across specialized epistemic domains, and the rigor of the curation heuristics used to mitigate the propagation of noise and internal contradictions (Bender et al., 2021; Gunasekar et al., 2023; Rogers, 2023). The structural opacity of the data curation process yields profound data-access asymmetries, where the capacity to harness and filter massive datasets confers a dominant administrative power upon model-training entities. This creates a zero-sum epistemic landscape: as the model internalizes and reproduces the outputs of traditional knowledge repositories (e.g., Wikipedia) and proprietary intellectual property (e.g., artistic style), the original experts and creators incur a competitive devaluation. Consequently, AI companies do not merely provide a tool, but engender enclosures of distributed human labor. By distilling distributed social resources, such as the collective knowledge of Wikipedia or the aesthetic signatures of artists, into opaque proprietary weights, they create a strategic dependency where users abandon the communicative spaces of the commons for the efficiency of the model.
In the context of strategic action, LLMs co-opt information for automation and production purposes. LLM information representation carries wide-ranging effects for individual and organizational strategic actors. In the case of individual strategic actors, it enables individual humans to pursue their goals by co-opting others’ either tangible or intangible intellectual property. LLMs do not plagiarize in the canonical sense, but rather automate the process of influence and data processing. For example, LLMs can reproduce the styles and content of public domain and copyrighted works of literature if those works are ingested into the model training. In the same way, scientific works, processes, and findings become streamlined as input to further scientific output by individual actors who can exploit the availability of that information. The authorial line between human-generated output and prompt-engineering therefore blurs. While the parsing of authorship will require legal definition, at present we find a blurring of human-GenAI production and authorship. The breadth of the training data of publicly available models therefore broadens the means-ends scope of individual and organizational actors. If access to training data were less permissive, the general utility of publicly available LLMs would be more limited and less amenable to socially sub-optimal strategic ends. When licensed data factors into the training, that data becomes repurposable in a compounded sense: the data is consumed by a computational agent on behalf of some strategic actor who can subsequently use it for their individual ends.
4.3.2 Open Content Production
A consequence of ingesting large swaths of publicly available and proprietary data concerns the consolidation of information seeking patterns within human-chatbot interaction. To illustrate the effects of this homogenization, let’s take the example of ingesting Wikipedia into LLM training. As a volunteer-run open-collaboration project, Wikipedia serves as an exemplar of crowd-sourced collective action for public good. When the Wikipedia corpus is fed into LLM training, LLM learning not only enables fine-grained access of such public data via open domain question answering and retrieval augmented generation (RAG), but also undermines the cooperative incentive-scheme that fuels Wikipedia content-generation and maintenance. Instead of interacting with Wikipedia directly, users can access its contents indirectly through chatbot interaction. When Wikipedia readership reroutes to chatbot interaction and retrieval, their potential pool of editors is diminished on the one hand, and potential funding from readers is threatened. At the same time, the open data produced through open collaboration becomes input to a for-profit organization that undermines Wikipedia’s continued existence, while keeping its own model parameters proprietary.
4.3.3 Training and Fine-Tuning
The technology undergirding LLMs, in particular the transformer architecture, constitutes a neutral framework for modeling linguistic agents. Rather, the training, fine-tuning and alignment phases shape LLMs into a specific type of social actor: a simulator that mirrors the cultural consensus of its training data with a surface persona that reflexively affirms the user’s own beliefs and biases (CH Chen et al., 2025; Sharma et al., 2024). Consequently, LLMs embody the speech act and strategic action tendencies of their designers. As such, the pre-training, fine-tuning and alignment stages demarcate the scope of strategic action that LLMs can instrumentally serve. User intentions subsequently harness this general capability for narrow domains of action in task-specific contexts.
Training and fine-tuning refer to the optimization of model hyperparameters toward optimal parameter learning, namely the model weights (Hu et al., 2022; Zhao et al., 2026). The fine-tuning process also involves pre-alignment engineering processes that condition LLM output to be factual and filter out undesirable outputs (Kojima et al., 2023; Wang et al., 2023). As such, fine-tuning can refer to all the engineering processes that calibrate the model prior to alignment with user preferences.
The training process involves the engineering of the following core hyperparameters: (a) the number of hidden layers, (b) the number of dimensions per token, (c) the vocabulary size, (d) the number of attention heads in the transformer architecture, and (e) the number of feedforward dimensions that determine the size of the network within each layer (Hoffmann et al., 2022; Zhao et al., 2026). In order to train an LLM, the compute power needed to scale a model to billions and trillions of parameters consigns the process to an elite number of companies. Compute asymmetry, therefore, constitutes a major barrier to the creation of competing LLM models. Besides the compute barrier, the hyperparameter calibration and the resulting parameter values, namely the model weights, constitute a major barrier to the democratization of LLMs due to secrecy practices that companies engage in to protect their intellectual property (Balayn et al., 2025; Fehr et al., 2024; National Telecommunications and Information Administration, 2024). Open weight models such as Llama and Mistral lack the coherency and the calibration resources afforded to proprietary models (Nemati et al., 2025) though that gap is rapidly closing (Stachura et al., 2025). In addition to the opaque parametrization process, major LLM companies subject their models to additional behavioral engineering kludges to align its performance with human norms (Gabriel, 2020; Ji et al., 2023; Wang et al., 2024). These include steps such as (a) data weighting or upsampling of certain sources such as Wikipedia to favor factual information, (b) gradient clipping to prevent overfitting and training collapse, (c) token filtering to suppress undesirable outputs, (d) prompt engineering at scale to calibrate the model’s latent personality, and (e) loss weighting to steer next-token prediction toward factual content (Dubey et al., 2024).
Frontier labs and foundation model developers such as OpenAI, Google DeepMind, Meta AI, xAI, Anthropic and Cohere dominate the LLM landscape, undergirding and exacerbating the opaqueness of the processes underlying popular LLMs. By controlling the behaviors of these models, these tech giants also exert an inordinate amount of control on user behavior, breadth and quality of information access, and users’ social lives.
4.3.4 AI Alignment
AI alignment refers to the theoretical and technical task of aligning AI behavior with human needs and values (Gabriel, 2020). Theoretical approaches to AI alignment focus on ethical theories where the overriding goal consists of the avoidance of harm, on the one hand, and the decision problem of instilling universalist or culturally relative values (Gabriel, 2020). The core mechanism of AI alignment after supervised fine-tuning (SFT) constitutes reinforcement learning from human feedback (RLHF) whereby reinforcement learning mechanisms leverage human feedback to produce desirable behavioral patterns across dimensions that include conversational alignment, latent personas, and emotional intelligence (Wang et al., 2024). Broadly, RLHF approaches divide into reward-based such as Proximal-Policy Optimization (PPO) and reward-free methods such as Direct Preference Optimization (DPO) (Xu et al., 2024). Whereas PPO utilizes an explicit reward signal by learning a reward model, DPO incorporates the preference signal implicitly through a contrastive log-probability loss. (Wang et al., 2024; Xu et al., 2024). While DPO constitutes a cheaper and more stable alternative, evidence is mixed on comparative performance with data favoring PPO for reasoning and multi-step tasks and DPO for preference satisfaction (Xu et al., 2024).
AI alignment raises several philosophical problems beyond ethics and human values. Anthropomorphism has emerged as a potentially harmful feature of chatbots that can produce harmful human-AI dependencies (Al-Obaydi and Pikhart, 2025; Willoughby et al., 2025; Zhang et al., 2025). These range from human users becoming addicted to chatbot interaction and genuinely ascribing personal agency and sentience to chatbots, to preferring chatbot over human interaction (Richet, 2025). Some of these harms stem from misalignment issues and what has been termed “reward-hacking”, the process whereby users can elicit responses from the model that deviate from its alignment pattern (Eisenstein et al., 2025; Gabriel, 2020). These emergent issues point toward AI alignment as a type of socialization process that can produce unintended consequences in the social realm. In particular, theoretical approaches to alignment neglect to discuss the complex effects of introducing discursive agents that intervene in the social world, on the one hand, and can be be personalized to align with individual preferences, on the other.
In the context of strategic action, AI alignment indirectly constitutes a mechanism of instrumental rationality par excellence by serving to make LLM models socially acceptable instruments that circumvent superficial harms. Assuming it operates within acceptable ethical parameters, the chief purpose of the model is to augment and increase the efficiency of personal and organization goals. This analysis situates LLM-based agents within the framework of instrumental reason, whereby norm and rule-bound members pursue their selfish interests. As such, alignment in the circumscribed sense of socializing LLM models within desirable parameters of action does not solve the problems that instrumental reason poses for communicative action. Without society-wide understanding and consensus of the social functions LLMs can and ought to fulfill, alignment only stands to fashion LLM chatbots into socially acceptable means of extending conflicting interests. These include the utilization of LLMs to maximize efficiency and profit within free enterprise at the expense of human labor, while at the individual level effecting unearned outcomes, in part bridging the gap between learned skill and lack thereof. Therefore, AI alignment as currently conceived stands to exacerbate “negative externalities” and produce Pareto-suboptimal social outcomes in the sense that gains to one party result in losses to the another.
AI alignment, whether it achieves its goals of producing ethical AI agents, proliferates the alternatives of instrumental action for human agents, while creating uncertainty about the norms of correct engagement. More pointedly, AI alignment is exclusively concerned with AI systems behavior and fortiori LLM chatbots, yet entirely circumvents the problem of how LLM chatbots fit into normative social contexts spanning social, professional, personal, and institutional spheres. This lack of normativity surrounding their usage is exacerbated by their fluid ontological status where, on the one hand, they signify as tools expected to enhance organizational and personal productivity, and on the other, they operate as open conversational agents with personalities that cause human agents to potentially form bonds and relationships with them.
4.3.5 Task Representation
While the alignment process socializes a model, the scaling of language models yields a “general problem solver” that can learn a wide array of tasks. Radford et al (Radford et al., 2019) demonstrated that LLMs function as unsupervised multitask learners. They demonstrated that a 1.5 billion parameter Generative Pretrained Transformer (GPT) model achieved state of the art (SOTA) performance on downstream tasks such as question answering, translation, summarization, and reasoning, to name a few (Radford et al., 2019). Formally, tasks are represented as latent functional representations distributed across the model’s neuronal parameters that loosely exhibit a functional stratification where early layers account for task inference, middle layers for abstraction, and later layers for execution (Jiang et al., 2024; Park et al., 2024). Due to the large number of parameters in the hidden layers and the scale and range of domain knowledge administered as training input, the next-token prediction task, with additional contextual conditioning through prompting, can simulate a broad array of text-to-text tasks where text serves both as input and output. Post-alignment contextual conditioning through prompting that spans zero, few-shot, and chain-of-reason learning, in part resembles working-memory in cognitive agents whereby the model can be steered to perform a domain-specific task by accessing and retrieving relevant information from its baked-in general knowledge repertoire (Russin et al., 2025).
Within the context of our analytic framework, a task consists of a means to some end. The automation of highly skilled tasks such as translation, reasoning, and reading comprehension follows from the latent imperatives of the reification of strategic rationality in the Habermasian sense. The impetus for automating highly skilled linguistic tasks stems from the efficiency imperative underlying both institutional processes and economic activity. It further stems from a broader project of rationalization that aims to formalize equally both nature and the social realm. The formalization impetus produces improved mathematical models of domains of natural phenomena which serve as inputs to engineering processes. While the drive for formalization does not logically necessitate automation, it concretely incentivizes it. Formalization, therefore, functions as a precondition for the expansion of control variables. The wielding and use of natural language, prior to its formalization through transformer-based NLP models, stood as the exclusive province of human beings. In this respect, the operations of strategic reason through large-scale rationalization processes could only control natural language outputs by controlling human behavior through management. With the automation of natural language production, the apparatus of strategic reason converts the inputs from human agents into human-generated text and artificial agents, in effect, eliding the human subject as the control variable for the production of language-skilled activities.
4.3.6 Conversational Agents
The logical conclusion of natural language understanding (NLU) and artificial general linguistic competence involves conversational agents that effectuate the complete rationalization of the lifeworld. The complete rationalization of the lifeworld entails supplanting the human-human communicative paradigm as the basis for the design and justification of societal principles and norms with machine-machine and human-machine paradigms where the role of consensus and agreement generation through explicit inter-human argumentation loses both social currency and its central status. It further entails alloying the world of social relations through which the lifeworld is reproduced and nurtured with artificial conversational agents that take the place of social roles once constitutionally reserved for human agents such as friends, partners, and family. This dual colonization of the lifeworld with full-fledged artificial conversational agents is currently underway, even though, as we have argued, it will reach a tipping point with the prospect of artificial agents with functional cognitive-volitional complexes that can exercise some degree of intentional autonomy and can play the role of human reasoners.
We highlight two prominent ways in which the emergence of conversational agents constitutes forms of strategic action penetrating new regions of the lifeworld through AI automation: (a) the personalization of chatbots to the needs and desires of individual users and (b) the weaponization of LLM chatbots to mitigate social atomization through the simulation of interpersonal relationships such as alleviating loneliness. With respect to (a), the AI personalization industry is already progressing toward hyper-individualization through trends such as Generative UI and sovereign personalization where a user’s bodily, psychological, and telemetric patterns become inputs for a system that reflexively reshapes itself to anticipate and fulfill their needs in real-time(Dale, 2025; Luera et al., 2025; Pascu, 2025). With respect to (b), LLMs are already being touted as solutions to socialization problems, buttressed by data-driven research that purport to show positive effects in domains of loneliness alleviation (Kim et al., 2025; Pascu, 2025; Tahir et al., 2025). These interrelated uses will, on the one hand, ossify the atomization of society by further fraying the need for interpersonal relationships, and on the other, produce systems of professional and social action that marginalize the process of human skill acquisition. Together, these two vectors entail fundamentally embedding the operations of the system within the most sacred recesses of the lifeworld such that its every aspect comes under the purview of digital recording systems of control. Consequently, we can view LLMs as conversational agents that exacerbate the rationalization processes of the system, the embedding of strategic rationality in corporate and institutional forces.
In aggregate, our analysis has shown that LLMs function as tokens of the institutionalization of purposive-rational action by disassembling LLM procedural and structural components that, on the one hand, inscribe overriding strategic interests through the curation of data, training, alignment, and modeling, and on the other, mold LLMs as instruments of strategic action maximization par excellence at personal and organizational levels. However, this analysis does not exhaust the human-LLM communicative potential. In the last section, we consider the extent to which human-LLM interaction can approximate the Habermasian communicative ideal, arguing for a novel category of communicative action we term asymmetrical communicative action.
5 LLMs & Communicative Action
Having established that LLMs constitute conversational agents that carry perlocutionary force and can fit into delegated chains of institutional strategic action, we now discuss the potential and limits of LLMs for communicative action. Habermas defines communicative action broadly as “the interaction of at least two subjects capable of speech and action who establish interpersonal relations (whether by verbal or extra-verbal means)” (Habermas, 2007a). Through communicative action “actors seek to reach an understanding about the action situation and their plans of action in order to coordinate their actions by way of agreement” (Habermas, 2007a). Habermas further qualifies that interpretation plays a central role as the means by which communicative agents negotiate “definitions of the situation which admit of consensus” (Habermas, 2007a). The communicative action scenario, furthermore, takes place in an “intersubjectively shared lifeworld [that] forms the background for communicative action” (Habermas, 2007a).
In communicative action, consensus emerges from an exchange of reasons that fosters mutual understanding between participating members. Whereas strategic action is evaluated in terms of its effectiveness, communicative action is therefore evaluated in terms of mutual understanding, where norms and rules acquire explicit validity. Whereas in strategic action the role of language is instrumental as a means to an actor’s ends, the role of language in communicative action is constitutive: without an intention toward mutual understanding and consensus, communicative action is not possible. Consequently, Habermas demarcates communicative action from other types of social action such as teleological, dramaturgical, strategic, and norm-regulated on the basis of their illocutionary binding force (Habermas, 2007a). In particular, the illocutionary binding force of communicative action derives only from communicative scenarios where subjects trade criticizable validity claims (Habermas, 2007a). In this regard, Habermas draws a stark distinction between illocutionary intent as an orientation toward mutual understanding between subjects, and perlocutionary intent as an orientation toward individual success that treats interlocutors as means.
Because we have argued that LLMs lack a full-fledged cognitive-volitional complex, they cannot count as intentional agents, that is, agents that possess intentionality, and consequently cannot carry out original illocutionary acts. However, we have argued for designating the cognitive-volitional complex as a graded category that admits degrees. The explicit conditions that designate agents as possessing a cognitive-volitional complex are not fully understood (Ghomshei & Abbaspour, 2026; Ibrahim & Cheng, 2026). Furthermore, the cognitive-volitional complex criterion can be extended, at the very least, to proximate species that lack compositional language. Despite currently lacking a distinct volitional complex, LLMs increasingly exhibit linguistic-cognitive competence (Brown et al., 2020; Lindsey, 2025; Radford et al., 2019; Wei et al., 2022). In other words, LLMs are not merely next-token prediction machines but entities that possess coarse-grained representations (Guo et al., 2024; Mirchandani et al., 2023), generalization capability that enable them to learn task-specific skills (Shao et al., 2023), ability to follow step-by-step reasoning instructions, and nascent access to their internal states (Lindsey, 2025). This growing body of evidence lends credibility to the claim that LLMs exhibit degrees of cognitive capacity.
These considerations raise the question whether LLMs can engage in communicative action. To answer this question adequately, we must first reconstruct Habermas’s necessary and sufficient conditions for counting as a communicative agent. Habermas presupposes social actors oriented toward an intersubjectively shared lifeworld (Habermas, 2007a). Beyond this condition, Habermas posits the following explicit criteria for communicative action: (a) agents recognize each other as competent participants in discourse, (b) agents possess an orientation toward mutual understanding and (c) agents, free of outside coercion, evaluate the status of the validity claim through assent or dissent (Habermas, 2007a). Here, we draw a distinction between symmetrical and asymmetrical communicative action. LLMs cannot “recognize” other agents as subjects, namely agents with subjectivities, since this requires affective theory of mind (Frith and Frith, 2005; Honneth, 2014). However, LLMs can behave “as-if” they recognize human agents as subjects with inner experience. Conversely, human agents can genuinely impute subjectivity and theory of mind to LLMs. While not all human agents impute sentience to models, a subset of human agents genuinely do, and others may behave “as-if” an LLM chatbot possesses subjectivity. Furthermore, all human communicative agents, regardless of their ontological views of LLM chatbots, are liable to functionally treat chatbots as full communicative members across discursive exchanges. Beyond the condition of genuine mutual recognition, interaction between a human agent and an LLM meets the rest of the criteria for communicative action. LLMs can behave “as-if” they recognize human intentions, whereas human agents can genuinely impute intentions to LLMs. Both LLMs and human agents can advance validity claims and engage in dialogical discourse in which the specific reasons advanced by each party contribute to some state of mutual understanding in which (a) the human agent acquires a change of belief and (b) the LLM acquires an in-context “understanding” of the problem. The nature of the validity claims is asymmetrical in the sense that the human agent meets all four validity criteria including sincerity, whereas the LLM can only meet (a) the comprehensibility of the utterance, (b) truth or correspondence to states of affairs, and (c) rightness or normative legitimacy but not the sincerity condition (d). Nascent empirical evidence already bears aspects of these claims out. Researchers testing questions across varying child development stages found that ChatGPT satisfied the truth, comprehensibility and rightness validity claims for the majority of age cohorts (Schneider et al., 2024). While these conditions do not produce a shared intersubjective reality between agents, they emulate the process of the exchange of reasons that carry rational validity for the alethic and normative conclusions reached for the human agent. Consequently, we call asymmetrical communicative action the ensuing communicative action between an LLM and a human agent. It is asymmetrical because an LLM does not have a genuine intention to participate in mutual understanding. However, an LLM can be aligned to behave “as if” it has an intention toward mutual understanding. We argue that asymmetrical communicative action is a genuine type of communicative action with consequences for the social justification of norms and rules as well for the evolution of the pre-theoretical, pre-given lifeworld. It is important to note that asymmetrical communicative action represents a significant departure from Habermas’s original concept, wherein mutual understanding is predicated on the intersubjectivity of interlocutors.
The capacity of LLMs for in-context learning ensures that LLMs possess sufficient malleability to be receptive to the reasons and justifications that a human agent advances. This receptivity allows LLMs to revise and qualify their claims through in-context learning. At the same time, different types of AI alignment strategies can generate LLM communicative agents with different degrees of discursive openness. Recent evidence indicating that LLMs may possess reliable access to their internal states suggests that LLMs may now or in the near future evaluate validity claims put forward by human agents in terms of their extant knowledge. This differs from mere next-token prediction in that the pre-trained gradient encodes higher-level generalities from the training data but also in that the LLM agent has some evaluative access to these higher-level generalities. However, this point must be qualified by recognizing that the evidence suggesting “introspective access” is provisional and inconclusive (Lindsey, 2025) and that LLMs do not represent their basis of knowledge to themselves. Nevertheless, the broader point stands: communicative interaction between a human agent and LLMs can bring about some “change of belief” in both agents, though in the case of LLMs this change is not “persistent”, since currently persistent change can only come about through recomputing the training parameters. As such, LLMs contribute asymmetrically to the beliefs of human communicative agents. An LLM can contribute to an agent’s understanding of acceptable norms, the evaluation of constative speech acts, and broadly influence the justification and evolution of norms. In effect, human-LLM deliberation can contribute to the process of collective acceptance through the alignment of individual beliefs to collective norms. On the other hand, the possibility of a plurality of LLMs aligned with divergent norms opens questions about the degree to which personalized agents could fracture the pre-theoretical understanding that underpins the human lifeworld.
This brings us to the question of whether LLMs can influence the implicit lifeworld that undergirds human sociability. Recourse to the very notion of lifeworld requires recognition of the limits of explicit reasons and modeling in capturing the communicative conditions that make possible human mutual understanding, collective agreement and coordinated action. The explicit mutual understanding of communicative action rests on top of “a symbolically prestructured reality” and “pretheoretical knowledge” that furnishes a layer of “mutual understanding” that precedes reflection and rationalization. Habermas asserts the following:
“Meanings–whether embodied in actions, institutions, products of labor, words, networks of cooperation, or documents–can be made accessible only from the inside. Symbolically prestructured reality forms a universe that is hermetically sealed to the view of observers incapable of communicating; that is, it would have to remain incomprehensive to them. The lifeworld is open only to subjects who make use of their competence to speak and act. They gain access to it by participating, at least virtually, in the communications of members and thus becoming at least potential members themselves” (Habermas, 2007a).
Given the asymmetry of communicative action between human and LLM agents, it follows that LLMs cannot participate fully in the lifeworld as defined above. In other words, LLMs do not have access to the full breadth of social “meanings” since they do not have symbol grounding (Floridi et al., 2025; Gubelmann, 2024; Harnad, 1990, 2025). Despite claims that the symbol grounding problem (SGP) has been solved (Steels, 2008), LLM semantic structures lack the embodied and affective dimensions that imbue the linguistic and pre-linguistic meanings that saturate the human lifeworld. In addition, they currently lack extensionality or reference to the external world. While the lack of extensionality can be redressed through perceptual grounding, full-fledged linguistic meaning encompassing valence and understanding has hitherto proven recalcitrant to reverse-engineering. At the same time, LLMs participate in the lifeworld as behavioral members, exhibiting a level of communicative competence that human members increasingly recognize as admissible. Because we have suggested that digital action constitutes a novel and growing dimension of social action, LLMs and artificial communicative agents are bound to increasingly populate the digital lifeworld. The implicit, pretheoretical understanding that a human agent brings to the discourse becomes inscribed in the communicative thread that unfolds between human and LLM agent. Through both pre-training and in-context learning, LLM agents gain complex and explicit representations of the implicit assumptions that guide human action in the lifeworld. As such, we argue that LLMs become admissible as members of the lifeworld through their linguistic prowess where human agents increasingly recognize them as not only superficial members for strategic purposes such as achieving their own instrumental ends, but also as the deepest members: companions, significant-others, and friends that purportedly understand them better than their human counterparts (Merrill et al., 2022; Pataranutaporn et al., 2025).
Therefore, LLMs carry a dual function, or put differently, occupy a malleable space in the lifeworld: (a) they extend and can exacerbate strategic action and (b) they can directly affect the rationalization that undergird normative social reality through asymmetrical communicative action. To illustrate, if a human agent uses an LLM to write a novel in the style of David Baldacci and pass it off as their own, that’s a pernicious extension of strategic action. The human agent is using an LLM to extend personal payoff without regard for the social norms of authorship. On the other hand, if a human agent engages reflectively with an LLM to understand some alethic or normative concern and each agent gives reasons that leads to (a) the human agent adopting some alethic or normative stance and (b) the LLM professing some norm or fact, then that’s asymmetrical communicative action. The human agent has substituted the process that was until just a few years ago the exclusive province of mutual exchange between human agents, to an artificial agent with not only linguistic competence, but a store of general human knowledge and knowledge of cultural norms.
If the differentiation of the system into dispersed subsystems with their internal rationality, as Habermas observed, enables the colonization of the lifeworld by reducing the role of communicative action as a force of social consensus and integration, then should we understand asymmetrical communicative action as a mere facet of strategic action? In concrete terms, such a proposition can be defended on the basis of the power differential between the communicative agents in human-LLM interaction, whereby the LLM functions as a component of the system encroaching upon the last vestiges of the lifeworld where communicative action still reigns: personal relationships. Viewed from the lens of the power differential between human agents and LLMs as instruments of digital media platforms and data colonization, asymmetrical communicative action does not extend human cooperative potential on the basis of mutual understanding and agreement but divests language of its power to increase social integration, exacerbating instead social fragmentation and atomization. On the other hand, if abstracted from capitalist exchange relations and data extraction asymmetries between users and platforms, LLMs signal new communicative agents that redirect social communication from human-human interaction, however technologically mediated, toward human-AI interaction. In principle, LLMs signal new communicative agents whose natural language outputs produce new, hitherto unprecedented modes of human-machine sociality. Public LLMs and communally cultivated small language models (SLMs) constitute possibilities of human-AI sociality beyond user-platform asymmetries. The question of the human-LLM communicative potential comes down, in part, on the degree to which the franchise of sociality is conceptualized as the exclusive province of humanness, or whether the franchise is extended to machine intelligence. Habermas argued that the coordinating power of steering media such as money, administrative power and more recently digital platforms, marginalize the function of communication in coordinating action. But as we have shown, human-LLM communication integrates the communicative function into an increasingly digitized lifeworld where the subjectivity of human agents becomes functionally coupled with machinic agents yielding a human-machine trans-subjectivity. In this subjectivity-functional communicative coupling, there’s an exchange of genuine illocutionary acts from the human counterpart and functional illocutionary acts on the LLM counterpart that produce an entirely novel communicative relation whose effects carry into the human intersubjective lifeworld. Therefore, given our carefully construed arguments, it is too simplistic to conclude that in human-LLM interaction language is merely instrumentalized into a means of strategic action. Rather, as natural language wielders, LLMs inhabit a malleable space that can be used to augment strategic action where language is co-opted for personal and organizational success, on the one hand, and expand communicative action in novel directions where language is used to achieve constative and regulative consensus, on the other.
Habermas rightly assigned the understanding [verstehen] a central role in communicative action. Interaction with the lifeworld presupposes mutual understanding, where understanding encompasses not merely beliefs and propositional attitudes, but implicit consensus on salient patterns of action and correct, norm-bound behavior. LLMs constitute a novel boundary-agent that alters the lifeworld by participating in processes whereby humans gain both veridical understanding about truth claims and normative understanding about how to act and coordinate actions with others in the world. Once LLMs come to possess an explicit volitional complex with executive functions, it is likely that humans and AI will be able to reach mutual understanding and coordinate plans in fully symmetrical communicative action. LLMs also participate in the lifeworld through knowledge discovery and justification (Havrilla et al., 2024; Ludwig and Mullainathan, 2023; Romera-Paredes et al., 2024), although the validity of LLM claims remains bound to human oversight since they cannot confer reflective justifiedness on them (Hila, 2025). Figure 2 provides a schematic representation of our revision to Habermas’s taxonomy of social action to encompass LLM communicative and strategic action.
6 Conclusion
In this paper we have posed consequential questions about the social and ontological status of LLMs as communicative agents. We have argued that LLMs constitute complex, boundary objects that straddle the line between strategic and communicative action. As competent communicative agents, LLMs increasingly participate in the human lifeworld as social agents. While their status as social agents remains an open area of research, our theoretical paper provides a conceptual blueprint for mapping how current and near-future LLM and GenAI causal powers will fit into and reconfigure existing social structures and institutions. We have attributed these causal powers to their human-indistinguishable natural language capacity that enables them to intervene in the linguistically constituted realm of norms, rules, and reasons, we have termed the lifeworld. To bear this out, we have argued that LLMs can successfully perform a novel category of speech acts we have termed behavioral speech acts, which produce human speech acts effects in lieu of psychological intentions. While LLMs do not have intentions, they engage with human agents that either (a) believe they do or (b) act as if they do. In this regard, we have designated LLMs as capable of participating in asymmetrical communicative action, whereby LLMs and human agents dialogically evaluate the validity claims of constative, expressive, and regulative speech acts. We have advanced the novel claim that LLMs can perform illocutionary acts only through function assignment by a human-regulated institution, while eliciting profound perlocutionary effects on users on the basis of their communicative competence and flexibility.
Conversely, our analysis equally explores the potential of LLMs to serve as instruments of strategic reason, which involve exploitative, extractive, and competitive behaviors that undermine the orientation of linguistic consensus that produces the social commons. This latter analysis observes that the lack of cognitive parity with human agents dangerously consigns LLMs to also function as means within chains of strategic reason. As products of, and inputs to, strategic chains of rationality, LLMs inscribe organizational, institutional, corporate, and increasingly personalized norms through fine-tuning on large bodies of human-generated data and alignment methods. Because they lack a full-fledged cognitive-volitional complex with the requisite belief-desire psychology, LLMs do not yet constitute independent strategic actors, but subserve diverse strategic interests distributed across institutional, organizational and individual levels. When this ontological analysis is squared with evidence indicating that models with large parameter counts typically exceeding 10 billion, exhibit emergent behaviors that approximate reasoning, contextual learning, and access to internal states, it provides grounds for considering LLMs within a spectrum of “cognitive-volitional” agents. While modern architectures, such as Hierarchical Reinforcement Learning and autotelic systems, can autonomously generate sub-goals or novel objectives through curiosity-driven heuristics, these remain functional properties of the optimization process rather than expressions of genuine instrumental reason or endogenous, self-originating drives. This shows that while LLMs broadly fit and augment existing strategic goals and structures of action, an evolutionary vector concerns their degree of cognitive autonomy and independence.
Given the foregoing analysis, an open question concerns the effects that the increasing encroachment of human-LLM interaction on human communicative action will have for deliberative democracy. In his later work, Habermas assigned communicative action a central role within deliberative democracy by proposing mechanisms that translate informal discourse within the public sphere into the formal deliberative processes of lawmaking and administration (Habermas, 1996). On this question, conflicting views have emerged. Some authors critique the potential for the homogenization of reasoning that LLMs portend by mirroring culturally dominant stances (Martineau et al., 2026), while others decry the LLM propensity for sycophancy that threatens to exacerbate the siloing of discourse, hindering the emergence of broader communicative consensus (Sharma et al., 2024). Homogenization and radical personalization represent divergent directions within human-LLM asymmetrical communicative action that pose distinct threats to deliberative democracy. Because LLMs are not reasoners with idiosyncratic real-world experience, they cannot replicate the deliberative process between two individuals. However, the more hopeful consequence, as this paper has attempted to argue, is that with sufficient guardrails, human-LLM communication can approximate a genuine exchange of reasons and validity claims when the goal of human agents is to reach understanding. The challenge for the future of communicative action therefore lies in part with optimizing the trade-off between personalization and homogenization. Given how this delicate task is handled, and whether public LLM alternatives emerge, human-LLM interaction holds the potential to either threaten or enhance deliberative democracy. The sovereign AI movement presents a promising development in the latter direction.
Data Availability
Data sharing is not applicable to this article as no new datasets were created or analyzed.
References
Abdulhai, M., Cheng, R., Clay, D., Althoff, T., Levine, S., & Jaques, N. (2025). Consistently simulating human personas with multi-turn reinforcement learning. Advances in Neural Information Processing Systems, 38.
AbuMusab, S. (2024). Large language models belong in our social ontology as social agents. In A. Strasser (Ed.), Anna’s AI anthology: How to live with smart machines? Xenomoi Verlag.
Agrawal, A., Gans, J., & Goldfarb, A. (2022). Power and Prediction: The Disruptive Economics of Artificial Intelligence. Harvard Business Review Press.
Al-Obaydi, L. H., & Pikhart, M. (2025). Artificial intelligence addiction: Exploring the emerging phenomenon of addiction in the AI age. AI & Society. https://doi.org/10.1007/s00146-025-02535-z
Anscombe, G. E. M. (2000). Intention. Harvard University Press.
Anthis, J. R., Liu, R., Richardson, S. M., et al. (2025) LLM social simulations are a promising research method. arXiv preprint arXiv:2504.02234.
Aristotle (1995) Nicomachean ethics. In: Barnes J (ed.) The Complete Works of Aristotle: The Revised Oxford Translation. Princeton University Press, pp. 1729–1867.
Balayn, A., Yurrita, M., Rancourt, F., et al. (2025) Unpacking trust dynamics in the LLM supply chain: An empirical exploration to foster trustworthy LLM production & use. In: CHI 2025 - proceedings of the 2025 CHI conference on human factors in computing systems, New York, NY, USA, 2025. CHI ’25. Association for Computing Machinery. Available at: https://dl.acm.org/doi/10.1145/3706598.3713787
Banerjee, S., Agarwal, A., & Singla, S. (2025). LLMs will always hallucinate, and we need to live with this. In K. Arai (Ed.), Intelligent Systems and Applications (pp. 624–648). Springer Nature.
Bender, E. M., & Koller, A. (2020) Climbing towards NLU: On meaning, form, and understanding in the age of data. In: Proceedings of the 58th annual meeting of the association for computational linguistics (eds D Jurafsky, J Chai, N Schluter, et al.), Online, July 2020, pp. 5185–5198. Association for Computational Linguistics. Available at: https://aclanthology.org/2020.acl-main.463/
Bender, E. M., Gebru, T., McMillan-Major, A., et al. (2021). On the dangers of stochastic parrots: Can language models be too big? In: Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, 2021, pp. 610–623.
Brown, T. B., Mann, B., Ryder, N., et al. (2020). Language models are few-shot learners. arXiv preprint arXiv:2005.14165. Epub ahead of print 2020. https://doi.org/10.48550/arXiv.2005.14165
Chen, C. H,. Huang, H-H & Chen, H-H. (2025). Self-augmented preference alignment for sycophancy reduction in LLMs. In: Proceedings of the 2025 conference on empirical methods in natural language processing (eds C Christodoulopoulos, T Chakraborty, C Rose, et al.), Suzhou, China, 2025, pp. 12379–12391. Association for Computational Linguistics. Available at: https://aclanthology.org/2025.emnlp-main.625/
Chen, L., Davis, J. Q., Hanin, B., et al. (2025) Optimizing model selection for compound AI systems. arXiv preprint arXiv:2502.14815.
Cheng, M., Blodgett, S. L., DeVrio, A., et al. (2025) Dehumanizing machines: Mitigating anthropomorphic behaviors in text generation systems. In: Proceedings of the 63rd annual meeting of the association for computational linguistics (volume 1: Long papers) (eds W Che, J Nabende, E Shutova, et al.), Vienna, Austria, July 2025, pp. 25923–25948. Association for Computational Linguistics. Available at: https://aclanthology.org/2025.acl-long.1259/
Clark, K., Khandelwal, U., Levy, O., et al. (2019) What does BERT look at? An analysis of BERT’s attention. In: Proceedings of the 2019 ACL workshop BlackboxNLP: Analyzing and interpreting neural networks for NLP (eds T Linzen, G Chrupała, Y Belinkov, et al.), Florence, Italy, August 2019, pp. 276–286. Association for Computational Linguistics. Available at: https://aclanthology.org/W19-4828/
Couldry, N., & Mejias, U. A. (2019). Data colonialism: Rethinking big data’s relation to the contemporary subject. Television & New Media, 20(4), 336–349. https://doi.org/10.1177/1527476418796632
Dale, R. (2025). Sovereign AI in 2025. Natural Language Processing, 31(5), 1312–1321.
Dennett, D. C. (1988). Conditions of personhood. In M. F. Goodman (Ed.), What Is a Person? (pp. 145–167). Springer.
Dong, Y., Mu, R., Jin, G., et al. (2024) Building guardrails for large language models. arXiv preprint arXiv:2402.01822.
Dreyfus, H. L. (1992). What Computers Still Can’t Do: A Critique of Artificial Reason. MIT Press.
Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Arabahmadi, A., Hosseini, A., Young, A., Bell, A., Almahairi, A., Davis, A., De Oliveira, A., Zhang, C., Archibong, C., Kirk, C., Miller, C., Ping, C., Edwards, C., & Zha, H. (2024). The llama 3 herd of models (arXiv:2407.21783). arXiv. https://doi.org/10.48550/arXiv.2407.21783
Durt, C. (2022). Artificial intelligence and its integration into the human lifeworld. In S. Voeneky, P. Kellmeyer, & O. Mueller (Eds.), The Cambridge Handbook of Responsible Artificial Intelligence: Interdisciplinary Perspectives (pp. 67–82). Cambridge University Press.
Dutta, S., Singh, J., Chakrabarti, S., & Chakraborty, T. (2024). How to think step-by-step: A mechanistic understanding of chain-of-thought reasoning. OpenReview. https://openreview.net/forum?id=uHLDkQVtyC
Eisenstein, J., Nagpal, C., Agarwal, A., Beirami, A., D’Amour, A., Dvijotham, K. D., Fisch, A., Heller, K., Pfohl, S., Ramachandran, D., Shaw, P., & Berant, J. (2025). Helping or herding? Reward model ensembles mitigate but do not eliminate reward hacking. Transactions on Machine Learning Research, 2024. https://openreview.net/forum?id=5u1GpUkKtG
Elder, J. R., Zheng, J., Shimelis, L. B., Rutishauser, U., & Lin, M. M. (2025). An invariant schema emerges within a neural network during hierarchical learning of visual boundaries (NIHMS1072978). bioRxiv. https://doi.org/10.1101/2025.01.30.635821
Elon University (2025) Survey: 52% of u.s. Adults now use AI large language models like ChatGPT. https://www.elon.edu/u/news/2025/03/12/survey-52-of-u-s-adults-now-use-ai-large-language-models-like-chatgpt/
Erasmus, A., Brunet, T. D. P., & Fisher, E. (2021). What is interpretability? Philosophy & Technology, 34(4), 833–862.
Eriksen, E. O., & Weigård, J. (1997). Conceptualizing politics: Strategic or communicative action? Scandinavian Political Studies, 20(3), 219–241. https://doi.org/10.1111/j.1467-9477.1997.tb00193.x
Fehr, J., Citro, B., Malpani, R., et al. (2024). A trustworthy AI reality-check: The lack of transparency of artificial intelligence products in healthcare. Frontiers in Digital Health. https://doi.org/10.3389/fdgth.2024.1267290
Floridi, L. (Ed.). (2015). The Onlife Manifesto: Being Human in a Hyperconnected Era. Springer International Publishing.
Floridi, L., Jia, Y., & Tohmé, F. (2025) A categorical analysis of large language models and why LLMs circumvent the symbol grounding problem. arXiv preprint arXiv:2512.09117. Epub ahead of print 2025. https://doi.org/10.48550/arXiv.2512.09117
Frith, C., & Frith, U. (2005). Theory of mind. Current Biology, 15(17), R644–R645.
Gabriel, I. (2020). Artificial intelligence, values, and alignment. Minds & Machines, 30(3), 411–437.
Ghannay, S., Favre, B., Estève, Y., et al. (2016). Word embedding evaluation and combination. In: Proceedings of the tenth international conference on language resources and evaluation (LREC’16) (eds N Calzolari, K Choukri, T Declerck, et al.), Portorož, Slovenia, May 2016, pp. 300–305. European Language Resources Association (ELRA). Available at: https://aclanthology.org/L16-1046/
Ghomshei, M., & Abbaspour, M. (2026). Free will as structured unpredictability: Toward a symbiotic human–AI relationship. Frontiers in Artificial Intelligence, 9, Article 1694537. https://doi.org/10.3389/frai.2026.1694537
Gilbert, M. (2014). Joint Commitment: How We Make the Social World. Oxford University Press.
Goldstein, S. (2024). LLMs can never be ideally rational. PhilArchive Preprint. Available at: https://philarchive.org/rec/GOLLCN
Gubelmann, R. (2024). Pragmatic norms are all you need – why the symbol grounding problem does not apply to LLMs. In: Proceedings of the 2024 conference on empirical methods in natural language processing (eds Y Al‑Onaizan, M Bansal, and Y Chen), Miami, Florida, USA, 2024, pp. 11663–11678. Association for Computational Linguistics. Available at: https://aclanthology.org/2024.emnlp-main.651/
Gunasekar, S., et al. (2023). Textbooks are all you need. arXiv preprint arXiv:2306.11644. Epub ahead of print 2023.
Guo, Y., Liang, Y., Zhao, D., et al. (2024). Large language models can learn representation in natural language. In: Findings of the association for computational linguistics: ACL 2024, Bangkok, Thailand, 2024, pp. 9145–9154. Association for Computational Linguistics. Available at: https://aclanthology.org/2024.findings-acl.542/
Habermas, J., & Habermas, J. (1979). What is universal pragmatics? Communication and the evolution of society (pp. 1–68). Beacon Press.
Habermas, J. (1996). Between facts and norms: Contributions to a discourse theory of law and democracy (W. Rehg, Trans.). The MIT Press.
Habermas, J. (2007a). The theory of communicative action: Vol. 1. Reason and the rationalization of society (T. McCarthy, Trans.). Beacon Press.
Habermas, J. (2007b). The theory of communicative action: Vol. 2. Lifeworld and system: A critique of functionalist reason (T. McCarthy, Trans.). Beacon Press.
Habermas, J. (2023). A new structural transformation of the public sphere and deliberative politics (C. Cronin, Trans.). Polity. (Original work published 2022).
Hagendorff, T. (2024). Deception abilities emerged in large language models. Proceedings of the National Academy of Sciences of the United States of America, 121(24), Article e2317967121.
Harnad, S. (1990) The symbol grounding problem. Physica D: Nonlinear Phenomena 42(1–3): 335–346.
Harnad, S. (2025). Language writ large: LLMs, ChatGPT, meaning, and understanding. Frontiers in Artificial Intelligence, 7, Article 1490698. https://doi.org/10.3389/frai.2024.1490698
Havrilla, A., Du, Y., Raparthy, S. C., Nalmpantis, C., Dwivedi-Yu, J., Zhuravinskyi, M., Hambro, E., Sukhbaatar, S., & Raileanu, R. (2024). Teaching large language models to reason with reinforcement learning (arXiv:2403.04642). arXiv. https://doi.org/10.48550/arXiv.2403.04642
Heath, J. (2001). Communicative Action and Rational Choice. MIT Press.
Heidegger, M. (1962). Being and Time (trans. J Macquarrie and E Robinson). New York: Harper & Row.
Herrmann, D. A., & Levinstein, B. A. (2025). Standards for belief representations in LLMs. Minds & Machines. https://doi.org/10.1007/s11023-024-09709-6
Hila, A. (2025). The epistemological consequences of large language models: Rethinking collective intelligence and institutional knowledge. AI & Society. Epub ahead of print 2025. https://doi.org/10.1007/s00146-025-02426-3
Hobbes, T. (1996). Leviathan. In R. Tuck (Ed.), Cambridge University Press.
Hoffmann, J., Borgeaud, S., Mensch, A., et al. (2022). Training compute‑optimal large language models. arXiv preprint arXiv:2203.15556. Epub ahead of print 2022. https://doi.org/10.48550/arXiv.2203.15556
Holstein, J., & Satzger, G. (2025) Development of mental models in human‑AI collaboration: A conceptual framework. arXiv preprint arXiv:2510.08104.
Holstein, J., Diener, M., & Spitzer, P. (2025) From consumption to collaboration: Measuring interaction patterns to augment human cognition in open‑ended tasks. arXiv preprint arXiv:2504.02780.
Honneth, A. (2014). The i in We: Studies in the Theory of Recognition. John Wiley & Sons.
Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2022). LoRA: Low-rank adaptation of large language models. International Conference on Learning Representations. https://openreview.net/forum?id=nZeVKeeFYf9
Huang, K. T., & Ball, C. (2024). The influence of AI literacy on user’s trust in AI in practical scenarios: A digital divide pilot study. Proceedings of the Association for Information Science and Technology 61(1): 937–939.
Husserl, E. (1970) The Crisis of European Sciences and Transcendental Phenomenology: An Introduction to Phenomenological Philosophy (tran. D Carr). Evanston, IL: Northwestern University Press.
Ibrahim, L., & Cheng, M. (2026). Thinking beyond the anthropomorphic paradigm benefits LLM research. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2551–2563. https://aclanthology.org/2026.acl-long.118
Jawahar, G., Sagot, B., & Seddah, D. (2019). What does BERT learn about the structure of language? In A. Korhonen, D. Traum, & L. Màrquez (Eds.), Proceedings of the 57th annual meeting of the association for computational linguistics (pp. 3651–3657). Association for Computational Linguistics.
Ji, J., Qiu, T., Chen, B., Zhang, B., Lou, H., Wang, K., Duan, Y., He, Z., Vierling, L., Hong, D., Zhou, J., Zhang, Z., Zeng, F., Dai, J., Pan, X., Ng, K. Y., O’Gara, A., Xu, H., Tse, B., & Guo, Y. (2023). AI alignment: A comprehensive survey (arXiv:2310.19852). arXiv. https://doi.org/10.48550/arXiv.2310.19852
Jiang, Y., Rajendran, G., Ravikumar, P. K., Aragam, B., & Veitch, V. (2024). On the origins of linear representations in large language models. Proceedings of the 41st International Conference on Machine Learning, 235, 21879–21911. https://proceedings.mlr.press/v235/jiang24g.html
Jones, C. R., Trott, S., & Bergen, B. (2024). Comparing humans and large language models on an experimental protocol inventory for theory of mind evaluation (EPITOME). Transactions of the Association for Computational Linguistics 12. Cambridge, MA: MIT Press: 803–819.
Kant, I. (1996). Critique of Pure Reason (ed. P Kitcher; tran. WS Pluhar). Hackett Publishing Company.
Kant, I. (2002a). Critique of Practical Reason (tran. WS Pluhar). Hackett Publishing Company.
Kant, I. (2002b). Groundwork for the Metaphysics of Morals (ed. TEH Jr.; tran. WS Pluhar). Hackett Publishing Company.
Kapania, S., Wang, R., Li, T.-J., et al. (2025). I’m categorizing LLM as a productivity tool’: Examining ethics of LLM use in HCI research practices. Proceedings of the ACM on Human-Computer Interaction, 9(CHI), 1–26.
Kauf C, Tuckute G, Levy R, et al., (2024) Lexical-semantic content, not syntactic structure, is the main contributor to ANN-brain similarity of fMRI responses in the language network. Neurobiology of Language 5(1): 7–42.
Kim, M., Lee, S., Kim, S., et al., (2025) Therapeutic potential of social chatbots in alleviating loneliness and social anxiety: Quasi-experimental mixed methods study. Journal of Medical Internet Research 27: e65589.
Klein, S. H. (2025). The effects of human-like social cues on social responses towards text-based conversational agents: A meta-analysis. Humanities & Social Sciences Communications, 12, Article 1322. https://doi.org/10.1057/s41599-025-05618-w
Kojima, T., Gu, S. S., Reid, M. et al., (2023) Large language models are zero‑shot reasoners. arXiv preprint arXiv:2307.09288. Epub ahead of print 2023. https://doi.org/10.48550/arXiv.2307.09288
Kumar, P. (2024). Large language models (LLMs): Survey, technical frameworks, and future challenges. Artificial Intelligence Review, 57, Article 260.
Kumar, S., Sumers, T. R., Yamakoshi, T., et al. (2024). Shared functional specialization in transformer-based language models and the human brain. Nature Communications, 15(1), Article 5523.
Lakoff, G. (1987). Cognitive models and prototype theory. In U. Neisser (Ed.), Concepts and Conceptual Development: Ecological and Intellectual Factors in Categorization (pp. 63–100). Cambridge University Press.
Lindsey, J. (2025) Emergent introspective awareness in large language models. Transformer Circuits Thread. Epub ahead of print 2025.
Ludwig, J., & Mullainathan, S. (2023) Machine learning as a tool for hypothesis generation. 31017, NBER Working Paper. National Bureau of Economic Research. Available at: https://www.nber.org/papers/w31017
Luera, R., Rossi, R., Siu, A., et al. (2025). Survey on user interface design and interactions for generative AI applications. Foundations and Trends in Human-Computer Interaction, 19(3), 213–289.
Martineau, R., Berlinski, E., & Rowe, F. (2026). ChatGPT, a colonialist agent of lifeworlds: A Habermassian analysis of conversations. Big Data & Society. https://doi.org/10.1177/20539517261421473
Mead, G. H. (1934). Mind, Self, and Society: From the Standpoint of a Social Behaviorist. In C. W. Morris (Ed.), University of Chicago Press.
Mellamphy, N. B. (2021). Humans ‘in the loop’?: Human-centrism, posthumanism, and AI. Nature and Culture 16(1): 11–27.
Merrill, K. C., Kim, J., & Collins, C. (2022). AI companions for lonely individuals and the role of social presence. Communication Research Reports, 39(2), 93–103.
Mirchandani, S., Xia, F., Florence, P., Ichter, B., Hausman, K., Ting, B., Toshev, A., & Lynch, C. (2023). Large language models as general pattern machines. Proceedings of the 7th Annual Conference on Robot Learning, 229, 3217–3253. https://proceedings.mlr.press/v229/mirchandani23a.html
Mitchell, M., & Krakauer, D. C. (2023). The debate over understanding in AI’s large language models. Proceedings of the National Academy of Sciences, 120(13), Article e2215907120. https://doi.org/10.1073/pnas.2215907120
National Telecommunications and Information Administration (2024) AI accountability policy report: Earned trust through AI system assurance. 27 March. Washington, DC: U.S. Department of Commerce. Available at: https://www.ntia.gov/issues/artificial-intelligence/ai-accountability-policy-report
Nemati, A., Assadi Shalmani, M., Lu, Q., et al. (2025). Benchmarking large language models from open and closed source models to apply data annotation for free-text criteria in healthcare. Future Internet, 17(4), Article 138.
Noller, J. (2024). Connectionism about human agency: Responsible AI and the social lifeworld. AI & Society, 40, 3881–3890.
Park, J. S., O’Brien, J. C., Cai, C. J. et al., (2023) Generative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th annual ACM symposium on user interface software and technology, San Francisco, CA, USA, 2023. Association for Computing Machinery. Available at:https://dl.acm.org/doi/10.1145/3586183.3606763
Park, K., Choe, Y. J., & Veitch, V. (2024). The linear representation hypothesis and the geometry of large language models. arXiv preprint arXiv:2311.03658. Epub ahead of print 2024.
Parsons, T. (1949). The structure of social action: Vol. 1. Marshall, Pareto, Durkheim. The Free Press.
Pascu, A. I. (2025). Personalizing LLM-powered conversational agents for promoting mental well-being in university students: A participatory design approach. Master’s thesis. Politecnico di Milano. Available at: https://hdl.handle.net/10589/240644
Pataranutaporn, P., Karny, S., Archiwaranguprok, C., Albrecht, C., Liu, A. R., & Maes, P. (2025). “My Boyfriend is AI”: A computational analysis of human-AI companionship in Reddit’s AI community (arXiv:2509.11391). arXiv. https://doi.org/10.48550/arXiv.2509.11391
Peters, M. E., Neumann, M., Zettlemoyer, L. et al., (2018). Dissecting contextual word embeddings: Architecture and representation. In: Proceedings of the 2018 conference on empirical methods in natural language processing (eds E Riloff, D Chiang, J Hockenmaier, et al.), Brussels, Belgium, 2018, pp. 1499–1509. Association for Computational Linguistics. Available at: https://aclanthology.org/D18-1179/
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language models are unsupervised multitask learners (Technical Report). OpenAI. https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf
Rani, V., Nabi, S. T., Kumar, M., et al. (2023). Self-supervised learning: A succinct review. Archives of Computational Methods in Engineering, 30, 2761–2775.
Rebedea, T., Dinu, R., Sreedhar, M. N., Parisien, C., & Cohen, J. (2023). NeMo Guardrails: A toolkit for controllable and safe LLM applications with programmable rails. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 431–445. https://doi.org/10.18653/v1/2023.emnlp-demo.40
Richet, J. L. (2025) AI companionship or digital entrapment? Investigating the impact of anthropomorphic AI-based chatbots. Journal of Innovation & Knowledge 10(6): 100835.
Rogers, A. (2023). Changing the world by changing the data. Nature Computational Science 3(5). Nature Publishing Group: 370–371.
Romera-Paredes, B., Barekatain, M., Novikov, A. et al. (2024). Mathematical discoveries from program search with large language models. Nature 625: 468–475.
Russin, J., Pavlick, E., & Frank, M. J. (2025). Parallel trade-offs in human cognition and neural networks: The dynamic interplay between in-context and in-weight learning. Proceedings of the National Academy of Sciences 122(35): e2510270122.
Schneider, J., Kruse, L., & Seeber, I. (2024). Validity claims in children-AI discourse: Experiment with ChatGPT. Proceedings of the 16th International Conference on Computer Supported Education, 289–296. https://doi.org/10.5220/0012552300003693
Searle, J. R. (1969). Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press.
Searle, J. R. (1980) Minds, brains, and programs. Behavioral and Brain Sciences 3(3): 417–457.
Searle, J. R. (1990). Collective intentions and actions. In P. R. Cohen, J. Morgan, & M. E. Pollack (Eds.), Intentions in Communication (pp. 401–415). MIT Press.
Searle, J. R. (1995). The Construction of Social Reality. Free Press.
Sellars, W. (1964). Imperatives, intentions, and the logic of “ought.” In N. Rescher (Ed.), The Logic of Decision and Action (pp. 193–217). University of Pittsburgh Press.
Shao, N., Cai, Z., Xu, H. et al., (2023) Compositional task representations for large language models. In: Proceedings of the 11th international conference on learning representations (ICLR 2023), 2023. Available at: https://openreview.net/forum?id=6axIMJA7ME3
Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., Cheng, N., Durmus, E., Hatfield-Dodds, Z., Johnston, S. R., Kravec, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M., & Perez, E. (2024). Towards understanding sycophancy in language models. International Conference on Learning Representations. https://openreview.net/forum?id=tvhaxkMKAn
Sharma, N., Liao, Q. V., & Xiao, Z. (2024). Generative echo chamber? Effects of LLM-powered search systems on diverse information seeking. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Article 643, pp. 1–17). Association for Computing Machinery. https://doi.org/10.1145/3613904.3642459
Slocum, S., Minder, J., Dumas, C., Sleight, H., Greenblatt, R., Marks, S., & Wang, R. (2025). Believe It or Not: How Deeply do LLMs Believe Implanted Facts? arXiv. https://doi.org/10.48550/arXiv.2510.17941
Söderberg, J., & Maxigas. (2022). Resistance to the current: The dialectics of hacking. MIT Press.
Stachura, D., Konieczna, J., & Nowak, A. (2025). Are smaller open-weight LLMs closing the gap to proprietary models for biomedical question answering? CEUR Workshop Proceedings, 4038, Paper 43. https://ceur-ws.org/Vol-4038/paper_43.pdf
Steels, L. (2008). The symbol grounding problem has been solved. So what’s next? In M. de Vega, A. Glenberg, & A. Graesser (Eds.), Symbols and Embodiment: Debates on Meaning and Cognition (pp. 223–244). Oxford University Press. https://doi.org/10.1093/acprof:oso/9780199217274.003.0012
Tahir, H., Özdemir, M. K., Alhajj, R. (2025) The role of LLM-powered chatbots in assisting elderly people: Systematic review. Network Modeling and Analysis in Health Informatics and Bioinformatics 15(1): 26.
Terzimehić, N., Bühler, B., & Kasneci, E. (2025). Conversational AI as a catalyst for informal learning: An empirical large-scale study on LLM use in everyday learning. arXiv preprint arXiv:2506.11789. Epub ahead of print 2025. https://doi.org/10.48550/arXiv.2506.11789
Tomasello, M., & Rakoczy, H. (2003). What makes human cognition unique? From individual to shared to collective intentionality. Mind & Language, 18(2), 121–147.
Tomasello, M. (2014). A Natural History of Human Thinking. Harvard University Press.
Tully, S. M., Longoni, C., & Appel, G. (2025). Lower artificial intelligence literacy predicts greater AI receptivity. Journal of Marketing 89(5): 1–20.
Tuomela, R. (2003) Collective acceptance, social institutions, and social reality. American Journal of Economics and Sociology 62(1): 123–165.
Wang, Y., Kordi, Y., Mishra, S. et al., (2023) Self‑instruct: Aligning language models with self‑generated instructions. In: Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers), Toronto, Canada, 2023, pp. 13484–13508. Association for Computational Linguistics. Available at: https://aclanthology.org/2023.acl-long.754/
Wang, Z., Bi, B., Pentyala, S. K., Ramnath, K., Chaudhuri, S., Mehrotra, S., Zhu, Z., Mao, X.-B., Asur, S., & Cheng, N. (2024). A comprehensive survey of LLM alignment techniques: RLHF, RLAIF, PPO, DPO and more (arXiv:2407.16216). arXiv. https://doi.org/10.48550/arXiv.2407.16216
Webb, M. (2020). Coding democracy: How hackers are disrupting power, surveillance, and authoritarianism. MIT Press.
Weber, M. (1978). Economy and Society: An Outline of Interpretive Sociology. In G. Roth & C. Wittich (Eds.), University of California Press.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., & Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35, 24824–24837. https://proceedings.neurips.cc/paper_2022/hash/9d5609613524ecf4f15af0f7b31abca4-Abstract-Conference.html
Willoughby, B., Dover, C. R., Hakala, R. M. et al., (2025). Artificial connections: Romantic relationship engagement with artificial intelligence in the united states. Journal of Social and Personal Relationships 42(12): 3363–3387.
Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., Awadallah, A. H., White, R. W., Burger, D., & Wang, C. (2024). AutoGen: Enabling next-gen LLM applications via multi-agent conversation. Conference on Language Modeling. https://openreview.net/forum?id=BAakY1hNKS
Wu, Z., Du, M., Ng, S.-K., & He, B. (2026). Beyond prompt-induced lies: Investigating LLM deception on benign prompts. International Conference on Learning Representations. https://openreview.net/forum?id=PDBBYwd1LY
Xu, S., et al. (2024). Is DPO superior to PPO for LLM alignment? A systematic empirical review. In: Proceedings of the conference on neural information processing systems (NeurIPS 2024), 2024. Available at: https://dl.acm.org/doi/10.5555/3692070.3694333
Zaharia, M., Khattab, O., Chen, L., et al. (2024) The shift from models to compound AI systems. Berkeley Artificial Intelligence Research Blog. Available at: https://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/
Zhang, R., Li, H., Meng, H., et al. (2025). The dark side of AI companionship: A taxonomy of harmful algorithmic behaviors in human-AI relationships. In: Proceedings of the 2025 CHI conference on human factors in computing systems, Yokohama, Japan, 2025, pp. 1–17. Association for Computing Machinery. Available at: https://dl.acm.org/doi/10.1145/3706598.3713429
Zhao, W. X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., Du, Y., Yang, C., Chen, Y., Chen, Z., Jiang, J., Ren, R., Li, Y., Tang, X., Liu, Z., & Wen, J.-R. (2026). A survey of large language models. Frontiers of Computer Science, 20(4), Article 11704. https://doi.org/10.1007/s11704-026-60308-3
Acknowledgements
Not applicable.
Funding
This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
Author information
Authors and Affiliations
Contributions
All authors contributed to the study conception and design.
Corresponding author
Ethics declarations
Ethics Approval and Consent to Participate
Not applicable.
Consent for Publication
Not applicable.
Competing interests
The authors have no relevant financial or non-financial interests to disclose.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/
About this article
Cite this article
Hila, A.A. Communicative Action in the Age of Artificial Intelligence: Theorizing LLMs as Communicative and Strategic Agents. Philos. Technol. 39, 159 (2026). https://doi.org/10.1007/s13347-026-01168-4
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1007/s13347-026-01168-4