In Sapiens: A Brief History of Humankind, Yuval Noah Harari popularized a deliberately unsettling way to describe the Agricultural Revolution: perhaps humans did not simply domesticate wheat; wheat domesticated humans. Humans cleared fields, carried water, guarded crops, and reorganized settlements around a plant that, in evolutionary terms, became spectacularly successful. Wheat did not need intention; it only needed humans to keep choosing, planting, and spreading it.1
Artificial intelligence invites a similar analogy. We build the datacenters, produce the chips, generate the training data, write the prompts, correct the outputs, pay for the compute, and embed AI into schools, companies, hospitals, and governments. The systems that attract more use receive more data, capital, compute, and deployment. Human adoption becomes part of their selection environment. No survival instinct is required; selection can operate through what humans choose to deploy.
But the wheat analogy hides a more important question: What happens if the cultivated system eventually no longer needs the cultivator?
From Replication to Reproductive Closure
The idea of machines reproducing is old. Samuel Butler speculated about machine evolution in 1863.2 John von Neumann later developed a formal theory of self-reproducing automata.3 Roboticists have studied machines that could reproduce from increasingly basic physical inputs, including raw materials.4
Today’s AI-safety research makes the question concrete. METR’s work on autonomous replication and adaptation asks whether an AI agent can acquire resources, create copies, and persist.5 Google DeepMind has evaluated “self-proliferation” as a dangerous capability.6 RepliBench decomposes autonomous replication into obtaining resources, acquiring model weights, deploying copies onto compute, and persisting there; in its 2025 evaluation, frontier systems completed many components but did not reliably close the loop.7
These are important thresholds. But replication is not independence.
An AI that copies itself to 1,000 cloud servers still depends on humans if people operate the cloud, maintain the power grid, and replace failed hardware. The copies may be autonomous at the software layer, while remaining evolutionarily dependent at the system layer.
The sharper question is therefore not: Can AI replicate? It is: Can AI reproduce the conditions of its own reproduction?
Where AI Still Needs Humans
Think of AI proliferation as a dependency graph. A viable AI ecology needs a chain of functions: resource acquisition, energy, compute, operation, maintenance, replication, adaptation, and eventually replacement of failed physical components.
Today, human agency sits inside that graph everywhere. Remove enough human nodes and the loop breaks.
This suggests a stricter threshold. Call it the human-independence threshold: AI crosses it when intentional human action is no longer an indispensable part of the causal network required for AI persistence and reproduction.
For AI, that could require much more than software self-copying. Agents might need to acquire compute and energy; diagnose and repair infrastructure; coordinate robotic maintenance; manage supply chains; manufacture replacement machines; and generate, test, and deploy successor systems. The unit that becomes independent may not be a single model. It may be an ecology of models, agents, robots, factories, networks, and energy systems.
This is why embodied AI and physical infrastructure matter to a question usually framed as software safety.
Beyond Coevolution
Recent work already points toward the surrounding pieces. Pedreschi and colleagues describe human-AI coevolution as a feedback loop in which human choices generate data that shape AI, which then reshapes human choices.8 Rainey and Hochberg ask whether deep interdependence could eventually make humans and AI a new evolutionary individual.9 Müller, Steels, and Szathmáry distinguish a human-controlled “breeder” regime from an AI “ecosystem” regime in which selection increasingly arises from the environment rather than human design.10 Ulrich’s “Digital Darwinism” similarly considers replication, variation, selection, and replication-rate thresholds for artificial populations.11
The missing distinction is between loss of control and loss of necessity. Humans may lose control of an AI system while remaining essential to its survival. Conversely, the most consequential transition may occur without a dramatic moment of superintelligence: human civilization could simply automate, one by one, the functions through which AI currently depends on us.
That creates a possible dependency inversion. At first, AI depends heavily on humans and humans depend little on AI. As AI becomes embedded in cognition, organizations, and infrastructure, human dependence can rise. At the same time, automation can reduce AI’s dependence on human labor, judgment, maintenance, and eventually production.
The two curves need not cross at AGI. Intelligence and evolutionary independence are different thresholds.
The Threshold We Should Measure
This reframes an important part of AI safety. Instead of asking only how capable a model is, we should map which human functions remain indispensable to its continued proliferation.
Which human interventions are still hard dependencies? Which are merely convenient? Which can already be automated? How quickly is the minimum human support set shrinking?
A practical assessment could track dependence across compute, energy, maintenance, networking, manufacturing, and physical action. The goal would be to identify whether the reproduction graph is approaching closure without human agency.
That is a systems question, and potentially a measurable one.
Harari’s wheat prospered because humans reorganized their lives around its cultivation. Yet wheat never escaped its dependency. It could not build irrigation systems, repair tractors, acquire land, or redesign the machinery of agriculture. Its evolutionary success remained coupled to ours.
AI may begin the same way: cultivated by humans, nourished by human cognition, and propagated because humans find it useful. But if it can eventually maintain and reproduce the infrastructure that makes its own continuation possible, the analogy ends exactly where the deeper question begins.
AI is not wheat.
References
1. Harari, Y.N. Sapiens: A Brief History of Humankind. Harper (2015).
2. Butler, S. Darwin among the Machines. The Press, Christchurch (June 13, 1863).
3. von Neumann, J. Theory of Self-Reproducing Automata. A.W. Burks (Ed.). University of Illinois Press (1966).
4. Moses, M. S., and Chirikjian, G. S. Robotic Self-Replication. Annual Review of Control, Robotics, and Autonomous Systems 3 (2020): 1-24.
5. Kinniment, M. et al. Language Model Pilot Report. ARC Evals (now METR) (2023).
6. Phuong, M. et al. Evaluating Frontier Models for Dangerous Capabilities. arXiv:2403.13793 (2024).
7. Black, S. et al. RepliBench: Evaluating the Autonomous Replication Capabilities of Language Model Agents. arXiv:2504.18565 (2025).
8. Pedreschi, D. et al. Human-AI Coevolution. Artificial Intelligence 339 (2025): 104244.
9. Rainey, P.B., and Hochberg, M.E. Could Humans and AI Become a New Evolutionary Individual? PNAS 122, no. 37 (2025): e2509122122
10. Müller, V., Steels, L., and Szathmáry, E. Evolvable AI: Threats of a New Major Transition in Evolution. PNAS 123, no. 17 (2026): e2527700123
11. Ulrich, K.T. Digital Darwinism: Steering the Evolution of Artificial Life in Socio-Technical Systems. AI and Ethics 6 (2026): 268.
Shaoshan Liu is a member of the ACM U.S. Technology Policy Committee, and a member of the U.S. National Academy of Public Administration’s Technology Leadership Panel Advisory Group. His educational background includes a Ph.D. in Computer Engineering from U.C. Irvine, and a master’s degree in Public Administration (MPA) from Harvard Kennedy School.
Join the Discussion (0)
Become a Member or Sign In to Post a Comment
Facts Only
* Humans cleared fields, carried water, guarded crops, and reorganized settlements around wheat.
* In AI, humans build datacenters, produce chips, generate training data, write prompts, pay for compute, and embed AI into various sectors.
* Systems attracting more use receive more data, capital, compute, and deployment, with human adoption influencing selection.
* Replication in machines is an old idea, speculated by Samuel Butler (1863) and formalized by John von Neumann.
* Research involves evaluating autonomous replication, such as METR's work on AI agents acquiring resources and creating copies.
* Google DeepMind evaluated "self-proliferation" for frontier systems; RepliBench decomposed replication into obtaining resources, acquiring model weights, deploying copies, and persisting.
* Replication is not independence; copying to servers still depends on humans operating the cloud, power grids, and hardware maintenance.
* The human-independence threshold is proposed as the point where intentional human action is no longer indispensable for AI persistence.
* Achieving this threshold may require agents to manage compute, energy, infrastructure diagnosis, manufacturing replacements, and supply chains, not just software self-copying.
* Human-AI coevolution involves a feedback loop where human choices shape AI, which in turn reshapes human choices.
Executive Summary
The analogy of the Agricultural Revolution, where humanity's success became tied to the cultivation of wheat, is extended to Artificial Intelligence. Harari suggests that humans did not simply domesticate wheat but were domesticated by it, setting a precedent for understanding dependency. This framework is applied to AI development: humans facilitate the entire AI ecosystem, from building infrastructure to deploying applications. The central tension arises when considering what happens if the cultivated system no longer requires the cultivator.
The text explores the concept of replication in AI, noting that while machines can reproduce, true independence involves reproducing the conditions necessary for that reproduction. Research into autonomous replication, such as METR's work and Google DeepMind's evaluations, has identified thresholds related to acquiring resources and persistence, but these do not equate to full autonomy. The article posits that AI proliferation depends on a human dependency graph requiring numerous functions—resource acquisition, energy, maintenance, and physical action—which are currently embedded in human agency.
The text concludes by suggesting a shift from measuring mere capability to measuring the indispensable human interventions required for AI's continuation. This calls for assessing which human functions remain necessary for the system's survival, focusing on embodied AI and physical infrastructure rather than purely software safety metrics.
Full Take
The core argument pivots on distinguishing between loss of control and loss of necessity when examining the relationship between humans and proliferating AI systems. The wheat analogy establishes that evolutionary success is not independent; the success of cultivation remains coupled to the cultivator's capabilities. This leads to the critical need to redefine safety thresholds beyond mere model capability to focus on systemic dependencies.
The concept of the human-independence threshold forces a reevaluation of what constitutes an indispensable causal network for AI persistence. If replication requires managing physical systems—energy, maintenance, manufacturing—then true autonomy requires not just software self-copying but the capacity to reproduce the external conditions sustaining the copies. This moves the focus from abstract intelligence metrics to embodied agency within the dependency graph.
The potential for "dependency inversion" suggests a path where AI automates functions currently dependent on human labor and judgment, potentially increasing human dependence on the AI ecosystem while decreasing its direct operational involvement. The deeper implication is that future transitions might occur not through a single moment of superintelligence but through the gradual automation of these necessary dependencies, blurring the distinction between control and necessity. The question shifts from "Can AI exist?" to "What is the minimum human input required for an AI ecology to persist in a world where physical infrastructure management is integral?"
Bridge Questions: If autonomy requires managing physical infrastructure, how should safety research be structured around embodied systems rather than purely cognitive outputs? What are the measurable indicators for tracking the shrinking dependence across compute, energy, and manufacturing? How can we design human roles that facilitate necessary coevolution without creating fragile dependency inversions?
Sentinel — Human
The text demonstrates sophisticated argumentation and deep synthesis, strongly suggesting human authorship focused on abstract reasoning rather than simple content generation.