Hiring is starting to be a conversation between different robots.
Most candidates now use AI somewhere in their application. Most employers of any scale now use AI somewhere in their screening. The middle of a recruitment process, the part where a candidate’s signal is supposed to meet a human evaluator’s judgement, has quietly become a conversation between two language models. The candidate’s career and the company’s hire both hang on what comes out the other side of that exchange.
That framing sounds dramatic. The research suggests it is not.
The candidate side
Our recent candidate survey, which formed the basis of the Talent Trust Gap report, found that two out of three candidates are now using AI somewhere across their application process. Cover letters, CVs, screening question answers, interview prep. Often all of the above.
This is not emerging behaviour anymore, it is the default. The “candidates are using AI” conversation has been running for two years, and most TA leaders accept the basic shape of it. What is worth saying again is that the function has largely not redesigned its assessment process around the new default. The forms, the screens, the take-homes, and the structured interviews most companies are running today were designed for a world that no longer exists.
The formats that have quietly lost the most signal are the ones TA functions still lean on hardest. The written cover letter is now almost entirely AI-mediated for many candidates. The free-text screening question (“why do you want to work here”, “tell us about a time when”) is barely more informative. The standard take-home task, the kind that takes a candidate two or three hours to complete, can often be finished in fifteen minutes with the right prompt. None of these have been formally retired by most companies. They are simply being run as if they still work.
That, on its own, would be enough to justify a serious rethink. But it is only half of what is going on.
The screening side
The newer and far less discussed half of the story sits on the employer’s side, and it is more uncomfortable than the candidate side.
In late 2025, a research team led by Jiannan Xu at the University of Maryland, with Gujie Li at the National University of Singapore and Jane Yi Jiang at Ohio State, published a paper titled “AI Self-preferencing in Algorithmic Hiring.” Using a large-scale controlled experiment with 2,245 human-written CVs and seven major language models, including GPT-4o, GPT-4-turbo, LLaMA 3.3, Mistral, Qwen and DeepSeek, they tested whether an LLM used as a screening tool would prefer CVs that the same model had also rewritten.
It did. Consistently. Across every model tested, with self-preference bias against human-written CVs ranging from 68 percent to 88 percent.
This is not a rounding error or a fragile finding. It is a structural bias that holds across the entire current landscape of commercial and open-source models. The mechanism the researchers identify is self-recognition. The model is implicitly detecting linguistic and stylistic patterns that match its own output, and weighting them positively. It is not that the AI-written CVs are better written, because the researchers controlled for content quality. The screener simply prefers content that looks like its own work.
The implication for hiring is direct. If a candidate uses ChatGPT to rewrite their CV, and your ATS or screening tool runs on an OpenAI model underneath, that candidate is significantly more likely to advance than a candidate of equivalent quality who wrote their own CV, or who used a different model. Which AI tools each side happens to use is now part of what shapes who gets hired.
To put the academic finding into a labour market context, the same research team simulated realistic hiring pipelines across 24 occupations and found measurable downstream effects on which candidates progressed and which did not. This is a new kind of bias, where the unfairness comes from which AI tool each side happens to use rather than from the candidate’s demographic profile. Current fairness frameworks were not built to detect it, and most of the AI hiring fairness conversation to date has been looking the other way.
There is a second-order effect worth flagging. As more candidates use AI to write their CVs, and as screening models reward AI-written content, the “winning” CV style increasingly converges on whatever the dominant model produces. Phrasing, structure, the rhythm of bullet points, the kinds of verbs used, all start to drift toward a single house style. Candidates who do not conform begin to look anomalous to the screener, regardless of their actual quality. The likely outcome over the next twelve to eighteen months is a rush to conformity, with CVs across the market converging on a narrow stylistic band optimised for whichever model is doing most of the screening. The candidates who can least afford that conformity, junior, non-native English, career-changers, are also the ones most likely to be punished for non-conformity.
Robots talking to robots
Put the two halves together and the shape of the system becomes clear.
Candidates increasingly write their applications with AI. Employers increasingly screen those applications with AI. The screening AI preferences inputs that look like its own outputs, which means it preferences AI-written inputs over human-written ones, and within AI-written inputs it preferences ones that match its own model family.
The candidate has no visibility into which model the employer is using. The employer has no reliable visibility into which model the candidate is using. The two systems negotiate the candidate’s onward progress between themselves, on the basis of stylistic resonance, before a human recruiter has read a word.
Consider how this plays out in a typical week. A scaling tech company is hiring a Senior Product Manager. The role attracts 400 applications in ten days. The first screen is run by an ATS-integrated LLM that scores CVs against the job description. Around 270 of those applicants used ChatGPT to refine their CV. About 60 used Claude. Maybe 70 wrote their own. The screening model, an OpenAI derivative, scores the 270 ChatGPT-refined CVs systematically higher than the equivalent-quality Claude-refined ones, and both groups higher again than the 70 human-written CVs. A shortlist of 40 goes to the in-house recruiter. The recruiter spends an hour reviewing those 40 with no idea that the funnel they are looking at has been shaped less by candidate quality than by candidate tooling.
This is what we mean by robots talking to robots. The top of the funnel for most high-volume hiring in 2026 is increasingly running on this loop, and the signal human evaluators are working from downstream has been pre-filtered by a system that is, demonstrably, biased toward AI-written content of a specific kind.
The implications go beyond fairness. They go to whether the hiring process is doing the job it is supposed to do. If the top of the funnel is no longer a reliable signal about the candidate, every downstream decision, the interview list, the panel debrief, the hire or no-hire call, is being made on a thinner evidence base than the people making those calls realise.
Most TA functions are not designed for this. Most ATS configurations are not designed for this. Most assessment workflows still treat the cover letter and the written screening response as if they carry signal about the candidate, when increasingly they carry signal about the candidate’s tooling.
What TA leaders should do about it
There are five positions worth taking. None of them are particularly comfortable.
1. Stop pretending you can detect AI use, and design as if you cannot.
AI detection tools are losing the arms race. False positives disproportionately punish non-native English speakers, and the tools that work today will not work in six months. Build the process on the assumption that AI was used somewhere in the application, and design assessment so that it is informative whether AI was used or not. Catching candidates out is a losing game. Designing assessment that survives AI use is a winnable one.
2. Find out what model your screening stack runs on, and decouple it from what candidates are using.
This is the most underrated position in the piece. If two out of three candidates are using ChatGPT, and your ATS or screening tool runs on an OpenAI model underneath, you have a self-preferencing loop sitting inside your hiring process. The mechanical advantage to candidates who happen to pick the same model family is real, measurable, and entirely invisible to the human reviewers downstream. Diversifying the model in your screening stack, or running a rules-based filter before any LLM evaluation, reduces that loop’s advantage. Most Heads of Talent cannot answer the question of what model their ATS actually uses. That should change this quarter.
3. Stop treating cover letters and write-in screening questions as signal.
At an industry level, this content is no longer a reliable indicator of the candidate. The information density of an AI-rewritten cover letter is approaching zero, and the information density of an AI-rewritten screening response is not far behind. Move the assessment burden onto formats that resist the loop. Live conversation, structured work product, paid trial tasks, and assessments specifically designed to evaluate how a candidate uses AI rather than whether they used it. The point is not to eliminate AI from the candidate’s side of the process, it is to stop relying on artefacts that AI has rendered uninformative.
4. Be explicit with candidates about what AI use is welcome and what is not.
The worst outcome is candidates using AI covertly because they assume it is banned. Anthropic’s published policy is a useful reference point. AI use to refine an application is welcome, AI use during live assessments and interviews is not. Pick a policy that matches the work, publish it on the careers site, and design the assessment around it. Ambiguity benefits no one. It quietly trains every candidate to game the system, and it punishes the ones honest enough to ask.
5. Put humans back into the decisions that matter, and accept the cost.
The mistake in 2026 is to assume that because AI saves time at the top of the funnel, the function can be run with fewer recruiter hours overall. The opposite is closer to the truth. When the top-of-funnel signal is degraded, the cost of that degradation is paid further down the process, by hiring managers and recruiters making judgement calls on weaker evidence. The compensating move is to expand human time on the decisions that actually matter, the late-stage interviews, the structured debriefs, the calibration meetings, and to accept that the hours saved at the top of the funnel need to be re-spent in the middle of it.
The bigger picture
The temptation is to respond to AI in hiring by adding more AI. Faster screening, smarter matching, AI interviewers, AI debrief notes. Some of those tools are genuinely useful and worth using. But more AI does not, on its own, address the underlying problem, which is that signal is becoming structurally unreliable in a system that is, increasingly, robots talking to robots.
The TA functions that handle this well in the next eighteen months will not be the ones that win the AI arms race on either side of the process. They will be the ones that are deliberate about where AI helps and where it quietly hurts, that are honest with candidates about what use is welcome and what is not, and that resist the false economy of letting humans out of decisions where human judgement is still, by some distance, the most reliable signal available.
The trust gap in modern hiring is not really about candidates. It is about whether the process itself can still be trusted to surface the people it is meant to surface. That is a question every Head of Talent should be sitting with in 2026, because the cost of getting the answer wrong compounds quickly, and it compounds in the kinds of hires that take twelve months to show up as a problem.
