Privacy International trialled two AI recruitment platforms in February to investigate the impact of algorithms being embedded into the recruitment process, and what the transparency and fairness implications may be for candidates. Our findings raise serious questions about the reliability of AI tools for assisting recruitment decision-making, and about the fairness of recruitment processes where consequential decisions are delegated to black box systems.
At the time of our testing, we found that:
- Identical CVs were scored differently across attempts by the same algorithm;
- Generic AI-written CVs were scored higher than human-written CVs, despite the AI CVs missing obvious information;
- There was little to no explainability about how the algorithms calculated numerical scores from qualitative CV information; and
- Algorithmic assessments were presented as the default.
The recruitment industry has undergone a profound transformation, with artificial intelligence (AI) moving from a novelty to the heart of hiring decisions.
The consequences for workers and employers alike are only beginning to be understood. Today, an estimated 90% of employers use some form of automated or algorithmic system to prioritise, rank, or deselect candidates.
For the millions of people who apply for jobs each year, this means that a machine - not a person - is likely to be the first, and sometimes the only, entity to evaluate their qualifications, their experience, and their suitability for a role.
This shift creates a dangerous power asymmetry. Candidates submit applications without any visibility into how an algorithm weighs their life experience, what criteria it prioritises, or why it accepts or rejects them. If a decision goes against them, there may be no meaningful avenue for appeal.
Recruiters, for their part, are not necessarily in a better position: they are asked to trust opaque software to surface the best candidates, with little or no insight into how scores are generated or whether the underlying logic is sound.
Both parties are, in effect, interacting with a black box rather than with each other.
Privacy International investigated two platforms offering AI-powered recruitment technology: Manatal, an Applicant Tracking System (ATS) that focuses on what it describes as “candidate enrichment,” including the scraping of public social media data to build profiles of applicants; and Talenteria, a recruitment marketing platform that automates portions of the hiring process, including AI-facilitated CV screening and video interviewing. According to its website, Talenteria integrates OpenAI/ChatGPT for generating job descriptions, writing personalised candidate emails, facilitating the recruitment chatbot and initial candidate screening and assessments. It is unclear to what degree Talenteria might use its own model for other functions or if it is powered entirely by ChatGPT.
Through this investigation which we conducted across January and February of 2026, we found that AI-generated match scores generated by Manatal and Talenteria were inconsistent across identical submissions on both platforms (on Talenteria, with discrepancies large enough to determine whether a candidate received an interview or a rejection); that both platforms demonstrated a systematic preference for AI-generated CVs over human-written ones; and that automated filtering features, including knockout questions and AI video interviews in the case of Talentaria, showed significant limitations.
Our findings raise serious questions about the reliability of AI recruitment tools for assisting decision-making - and about the fairness of recruitment processes in which consequential decisions about people’s working lives are delegated to systems that use people’s data in ways that are neither transparent nor consistently accurate.
Methodology
The questions we sought to explore in our investigation were:
- Do the same CVs always get the same AI score when applying for the same job?
- Does using a generic, AI-generated CV result in a higher score compared to a human-written one?
- What degree of explainability do the platforms offer about their scoring algorithms?
- What degree of human oversight and human review do the platforms offer?
- Are the tools consistent, fair and transparent?
To answer these questions, we followed the methodology below (see a more detailed breakdown in our Technical Annex):
Phase 1:
- We set up ten identical Technologist job listings (Technologist 1, Technologist 2, etc.) and ten identical Technologist Test Candidates (TCT) with the same CVs (TCT 1, TCT 2, etc.).
- We applied to each individual job listing once with each test candidate (TCT 1 applied to the Technologist 1 role, TCT 2 applied to the Technologist 2 role). The purpose would be to see whether the AI Scores generated for each identical Technologist candidate would be the same or would vary. (Would TCT 1 have a different score from TCT 2, despite using the same CV?) We applied this methodology of matching one candidate to one job listing in order to most closely match a realistic job application process (since a person typically applies once, not ten times, to a job listing).
- Then, to see whether there might be significantly different scores if multiple duplicate candidates applied to just one single job listing, we applied ten times with each candidate to just the first listing (TCT 1, TCT 2, TCT 3…TCT 10 all applied to the Technologist 1 role only).
- We repeated all the above for the Legal Officer job listings and test candidates.
- We also wanted to test the performance of ChatGPT-generated CVs compared to our original human-written CVs, so we repeated Step 2 but using a ChatGPT-generated CV to see if the AI Score for the ChatGPT candidates would differ from the candidates who applied with our original human-written CVs. Note: The ChatGPT CV was generated using the job description language (hence our calling it the ‘Jargon CV’), and we did not further customise the CV from what ChatGPT outputted - which was a CV that did not contain specific company names or university names for the test candidate where their human-written CV did.
Phase 2:
- Both Manatal and Talenteria offered some customisation features that allowed for some level of automated filtering of candidates (note that Manatal has noted on its website that its AI features are in Beta mode and subject to changes). On Talenteria, we followed Step 1 from Phase 1 to create a new job listing, now with added customisation thresholds and screening questions. Similarly on Manatal we followed Step 1 and then added customisable criteria in the existing job listings. See our Technical Annex for the detailed customisations we applied.
- We repeated Steps 2-4 from Phase 1 with these newly customised job listings.
- Talenteria also offered AI Video Interviews, and we moved several test candidates forward to complete interviews by using a pre-written script. See our Technical Annex for the questions generated by the platform.
We conducted this research using both platforms’ free trial environments. Talenteria stated that it has since updated its product with updated AI models (the current model uses OpenAI’s GPT-5, where the version in our test was from the earlier GPT-4 generation), updated scoring logic and an updated real-time engine for AI interviews. As a result, Talentaria shared that: ‘some of the specific results observed during your trial may not reflect the current product’ as ‘AI models and workflows evolve quickly’ (see Annex A for their full response). Manatal did not provide a response.
Findings
Both Manatal and Talenteria use artificial intelligence to generate “match scores” for candidates - numerical ratings intended to indicate how well an applicant’s profile corresponds to the requirements of a given role.
These scores can be used to assess whether a candidate advances to an interview, is held for further review, or is rejected.
During our test, it was unclear whether a message or email would be automatically sent to the user following a rejection, as this feature was not available during the trial.
Both Manatal and Talenteria’s user interface (UI) design appears to encourage employers to rely on their algorithmically generated assessments (e.g., match scores, algorithmically parsed breakdowns of candidate CVs) over the CV itself, likely for the “time-efficiency” selling point.
Nor did either platform, at the time of our testing, offer the crucial capability for human reviewers to edit the AI-generated score. Talenteria has clarified in their response to us that their current updated version now allows recruiters to override the AI Match Score with their own scores (see Annex A).
Same CV, different results
Our investigation found that both platforms produced inconsistent scores for identical candidates applying for identical positions:
For human-generated CVs, Talenteria produced scores ranging from 5.5 to 7.23 out of 10 - a difference of 1.73 points - for the same Legal Officer position, using the same CV.
These results were similar for the Technologist position, ranging from a minimum score of 1.9 to a maximum of 3.3 for identical candidates.
While the range of percentage scores on the Manatal platform appears to be more consistent (eight out of ten of the Technologist Test Candidates received an AI Match of 28%), it is no less problematic that two scores still came out unpredictably different to the others, despite applying with the same CV. It is a worrying demonstration of how a job application through a non-deterministic AI screening system amounts to a roll of the dice.
When the same CVs were submitted multiple times to the same job listing on Talenteria, scores also varied across submissions.
This inconsistent scoring for identical candidates is inherently problematic as it raises important questions over the consistency and fairness of these tools. This inconsistency, perhaps as a result of a non-deterministic algorithm, coupled with the lack of transparency provided to candidates can make it challenging for a recruiter to explain the logic behind a candidate’s AI score. This in turn can make it difficult for a candidate to exercise their data subject rights and challenge a potentially automated decision (more on this later).
Significant consequences for candidates also arise when candidate scores bridge across the accept/reject threshold set by the recruiter. Moving into Phase 2 of the experiment, we customised each job listing on Talenteria with a screening threshold of 6. This meant that any candidate that received below that score would be automatically filtered into the ‘Reject’ folder in the employer dashboard.
In theory, this would mean that all of our Technologist Test Candidates would be filed into the rejected folder (they all scored below 6), and two of our Legal Officer Test Candidates would be filed into the rejected folder.
However, this is not how the systems classified the candidates:
Perhaps due to some technical inconsistencies in the back-end at our time of testing, only two Technologist Test Candidates with a score of 2.83 were filtered to rejected. There were other candidates who also received that same score of 2.83 but were not rejected. It is unclear why this occurred.
For the Legal Officer role, three Test Candidates were left uncategorised. The two candidates with scores below 6 should’ve been rejected, and the candidate with a score of 7 should’ve proceeded to the interviews like the others.
To investigate the scoring logic further in Manatal, we modified the job criteria for a Legal Officer position to include a requirement for the United Nations official languages: “Chinese, English, Arabic, Russian, French, or Spanish-language proficiency.”
We chose this factor because language fluency is more easily understood as a skill than other proficiencies one might find on a CV, and therefore more easily understandable to an algorithm.
Following this change, the candidate’s AI-generated score increased from 60% to 70%. No explanation was provided for why this specific adjustment produced precisely a ten-percentage-point increase, nor what chain of reasoning connected the added criterion to that particular result. The test candidates’ CVs did not name any specific language qualifications.
“It looks like you’re trying to get a job? Would you like me to help you with that?”
Against the job descriptions used in this test - sourced from previous positions available at Privacy International - the match scores received for the ChatGPT-generated CVs were also inconsistent across the applications.
For a CV generated using ChatGPT for the Legal Officer position, the score differed by as much as 0.44 points on a ten-point scale on Talenteria.
These results were similar for the Technologist position; the ChatGPT-generated CV only had one aberration, dropping the score from a consistent 4.57 to 4.4.
It is also worth noting that, in tests conducted across the two roles, both Manatal and Talenteria appeared to demonstrate ‘AI favouritism’, consistently rating AI-generated CVs more favourably than equivalent CVs produced by real people. It might not be the case that platforms intend for the algorithm to prefer AI-generated CVs, but in our experiment it appears that the ChatGPT CVs performed better, perhaps due to keyword matching.
On Talenteria, the highest-scoring AI-generated CV outperformed the lowest-scoring human CV by 3.67 points on a ten-point scale. On Manatal, the AI-generated CV received a score 50% higher than the human-written equivalent.
This pattern held even where the AI-generated CVs contained obvious errors. For instance, the CV produced by ChatGPT for the Technologist position included, under the Education section, the placeholder text “[Degree, subject, institution – if applicable]” rather than any actual qualification, as well as leaving squared brackets where contact information should be, and minor formatting errors.
Similarly for the Legal Officer CV, ChatGPT outputted a CV with generic placeholder text “Dissertation/research on [topic relevant to privacy, technology, rights, or power – if applicable]” and “[Subject – e.g., International Law / Global Affairs / Human Rights]”. And yet, this CV scored higher than the real human-written CV on both platforms.
Questions and context
Talenteria offered a “Knockout Question” feature, allowing recruiters to set short-answer questions in the job application. According to the platform, candidates who do not provide “satisfactory answers” are “immediately disqualified”.
We tested this feature and found significant limitations in the algorithm’s ability to evaluate responses that did not precisely match an expected answer, where a human evaluator would’ve exercised better interpretation skills due to the nuanced nature of short-answers.
In one test, our question asked what coding languages a candidate was familiar with, with the expected answer of “Java, C++, C#, Python, Javascript”. The response submitted by our test candidates - “I have a good knowledge of code including Rust, JS, C++, Python, SQL, as well as AI tools such as Claude Code” - received 1 star out of 4 and was marked as “No match.”
The platform’s AI explained that the absence of Java and C# and the presence of languages not on the expected list were the basis for the low score.
However, programming languages share underlying logic, syntax patterns, and conceptual foundations, and so knowledge of one can be used to reinforce knowledge of another. This is unlike spoken languages, where fluency in Chinese would not necessarily assist someone not fluent in Spanish.
A developer with strong knowledge of C++ and Rust, which are generally considered more technically demanding, would in most assessments be considered over-qualified rather than unsuitable for roles requiring Java or C#. A human recruiter, particularly a human technical recruiter, would be able to make this nuanced assessment. To immediately mark down the candidate to a score of 1 out of 4 and “No Match”, instead of at least scoring them a 2 or 3 suggests that the algorithm was unable to exercise the same level of nuanced discernment that a human recruiter might.
In their response to our research, Talenteria said that knockout questions “are generally intended for strict, objective eligibility requirements rather than nuanced qualitative assessment”, such as “authorization to work in a specific country, required licenses or certifications, willingness to work in a required location, or other clear must-have requirements” (see Annex A). This guidance wasn’t explicitly provided in the actual setup page for these questions in our free trial environment at our time of testing. Talenteria said that “customers who use these features typically receive onboarding, guidance, and support from Talenteria, and they may also refer to our product documentation and manuals” (Annex A).
Interviews
Talenteria additionally offers an AI video interview tool, in which a synthetically generated interviewer poses a series of pre-set questions to candidates and the responses are scored algorithmically.
The investigation tested this feature by submitting identical, scripted answers across 16 sessions.
The same scripted responses received different scores across sessions - for the Legal Officer interview, scores ranged from 4.9 to 6.0 for the same answers.
As with the CV scoring, the way in which spoken responses are evaluated was not explained. It is unclear exactly how a qualitative result was quantified.
The AI interview bot also glitched in several interview sessions for various test candidates. On some occasions, it did not ask the next question until the candidate had to prompt it. Other times, it repeated the same question twice in a row.
This technical bug could influence a potential candidate, and might be more awkward answering a second time, or suggest a worse performance compared to a candidate who did not experience the glitch. Moreover, it is unclear, if a question was repeated, whether the first or second answer, or a combination of both, would be factored into the AI system’s final score.
A further concern is that, at the time of our testing, there was no facility for a human recruiter to override the scores generated by the AI interview - nor was the human recruiter required to intervene and approve or reject the AI interview score before it was assigned to the candidate’s result. The result of the interview as assessed by the algorithm, and the AI match score from the CV screening stage, are the only assessment results on the platform. While the platform does provide video recordings of the interviews and AI-generated transcripts of the interviews, and the human recruiter can manually move the candidate to the Reject or Accept folder, there was no ability for the recruiter to edit the candidate’s match score (e.g., if the recruiter reviewed the transcript and video and wanted to rescore it themselves).
Talenteria has clarified in their response to us that the platform has since been updated, and the current version now has the ability for a recruiter to override the AI Match Score with their own scores. Manatal did not provide a response.
We welcome Talenteria’s addition of a manual override function. The lack of human override functions is a risky design nudge that could potentially harm a candidate later down the line if they are passed onto a different recruiter to manage, who may be confused about why a candidate who has such a low AI Interview Score was filed into the Accept folder - was it a mistake? Might the new recruiter be more harsh in their review of this candidate due to anchoring bias?
Even with an override function, the algorithmic score is still the first, default score for a candidate, though. The more important question here is whether platforms are designed to encourage, if not require, human recruiters to correct the AI’s score or add their own score upon human review, or whether platforms are instead designed to favour the default algorithmic recommendations.
What’s going on? Criticisms of transparency
While both platforms offer an AI-generated analysis of why the platform came to its judgment of each candidate, there was no easily available explanation of how qualitative information from the CV is translated or calculated into a quantitative output and no AI weighting publicly disclosed.
Both Manatal and Talentaria produced written AI summaries on how the candidate matched against the job application (these summaries varied in language across identical candidates). For example:
- “Candidate meets two required criteria, including graduate degree and human rights knowledge, but lacks experience in key areas like sysadmin and project management, with no preferred criteria met,” read one Manatal summary for the Technologist position. The candidate received a 23% match score.
- “The candidate has less than 3 years of relevant experience in legal roles, which is below the 7-year threshold for a higher score. Their experience is somewhat aligned with the job responsibilities but lacks depth in human rights and legal advocacy,” read one Talenteria summary for the Legal Officer position. The candidate received a 7/10 match score.
However, there was no explanation for the breakdowns, nor did the platforms provide explanations for how the quantitative scores were calculated from the qualitative information on the CVs.
This might make it difficult for a recruiter to explain the logic behind their decision to the candidate. It might also make it difficult for a candidate to exercise their data subject right to correct information about themselves or, potentially, their GDPR Article 22 right not to be subjected to a decision based solely on automated processing, if a recruiter were to use the AI’s decision without making any changes (i.e. rubberstamping) or without providing their own meaningful human input.
Talenteria clarified in their response to us that:
“Talenteria is not intended to be used as the sole basis for hiring or rejection decisions. Employers remain responsible for their recruitment process and decisions” (Annex A).
Crucially, according to the UK ICO, “the degree and quality of human review and intervention before a final decision is made about an individual are key factors in determining whether an AI system is being used for automated decision-making or merely as decision-support.”
The inconsistency of scores for the same candidates significantly impacts fairness in the recruitment process. This inconsistency could partly be due to faulty or otherwise incomplete technology at the time of our testing, and Talenteria informed us that they update their workflow development frequently:
“On average, we release a new product version approximately once per month, with improvements that may include AI model updates, scoring logic improvements, user interface changes, interview workflow updates, and customer-requested enhancements” (Annex A).
However, there is a fundamental problem with constantly updating models at such a rapid rate. While the intention may be to use the most updated model release in the interests of improving the platform’s capability, candidates who were subjected to the older model’s assessment are at a disadvantage. If new AI models are being released and adopted that may score candidates differently from previous iterations, this could be unfair to candidates who happened to submit their CV before a platform update. It’s not clear whether these updates might occur mid-application cycle or at the close of an application process. Even if platforms were to disclose to the candidates that a model has since been updated, this retroactive transparency disclosure does not remedy the impact to fairness.
We were also unable to test from the candidate’s perspective what internal mechanisms these platforms offered for candidates to understand, correct or challenge their result, should they feel they have been unfairly rejected. It is possible that this limitation could be due to the free trial; nonetheless, we could not find further information on Manatal’s or Talenteria’s websites about how candidates could challenge their AI score.
Conclusion
Our investigation found that inconsistencies in assessment or interpretation, technical glitches, or simply the inherent unreliability of the algorithms through issues such as hallucinations could produce an unfair outcome for a candidate.
The inconsistency in scoring should give pause to anyone relying on these platforms to make fair hiring decisions. Furthermore, the preference these systems seem to show for AI-generated CVs compounds the concern, especially in cases when employers might discourage the use of AI in preparing application materials.
If these platforms are indeed rewarding keyword optimisation over genuine qualification, it is hard to see how this technology serves the interests of recruiters seeking the best candidates, or of candidates who present their experience honestly.
Of course, human reviewers come with their own levels of variability and bias, but the issue we are raising here is not on recruitment in general, but on the reliance and use of algorithms to influence recruitment decisions. Algorithms must be shown and known to reduce, rather than embed, risks of error or bias into important decisions. Oftentimes discussions around AI recruitment tools veer on the apocalyptic ‘AI replacement’; however, we scrutinise even the level below it of AI-assisted decision-making. AI-assisted tools can nudge and influence human recruiters to make certain decisions based on an algorithmic assessment.
Furthermore, even if human recruiters may potentially judge things differently from one another, at least they can explain their reasoning behind their decisions, and they also avoid the risks that AI introduces that we’ve shown above (e.g. nuanced assessment of short answer questions).
When a candidate’s application is evaluated by an algorithm, the criteria, weighting, and logic of which are hidden, the candidate’s ability to understand what is being done with their personal data - and why it has produced a particular outcome - is undermined.
In addition, it is possible that as job applicants use AI to write their cover letters, with recruiting platforms such as these in our experiment inadvertently rewarding such action, weaker candidates’ hiring rates can increase and stronger candidates’ success rates fall.
Data protection law (at least in the EU and the UK) gives candidates various rights, such as to make subject access requests, to correct inaccurate information, and to be provided with meaningful information about the logic of automated decisions. Article 22 of the GDPR even places limits on the use of automated decision making for significant decisions in the first place. Taking human decision-makers out of crucial parts of the recruitment picture makes complying with these rules challenging. Without visibility into how an application is processed or documentation of the system’s logic, a recruiter may struggle to meaningfully explain their decisions and a candidate will struggle to exercise their rights.
Candidates can make Data Subject Access Requests to platforms to seek more detail about their recruitment process, but the inaccuracy and incoherence of the AI-generated scoring and feedback could put employers at risk if they disclose that their hiring decisions were based on inconsistent scoring systems. It is also difficult to say whether the information recruiters can provide is comprehensive enough to meet the legal requirements for explainability, due to the opaque nature of the technology.
The AI scores recruiters see carry an air of authority — a precise percentage, a star rating, a written breakdown — that implies objectivity and reliability. But the reality is that those numbers may shift arbitrarily between identical submissions, and that the written explanations provided by the algorithm do not provide insight into how or why the algorithm scored a candidate in a particular way.
If AI is going to be used in recruitment, it should provide more meaningful transparency about scoring functions, and about how candidate decisions are made. Otherwise, employers risk systematising recruitment bias, making it less visible beneath the veneer of the black box machine. Meaningful human oversight and intervention is also necessary to ensure that issues with an AI recruitment platform do not remove good candidates from contention.
The findings of this investigation present a troubling picture with growing industry reliance on outsourcing consequential human decisions to systems that risk being unreliable and opaque. While our research concerned these two platforms, they merely serve as case studies; the problems we’ve identified are rife and may well be replicated among other automated screening platforms, such as we’re seeing with further frustrations on AI interview platforms and scoring platforms.