AI in recruiting is no longer something only employers use. Candidates now write resumes and cover letters with chatbots, and some use them on personality and ability tests. The research shows that applicants still distrust being judged by machines, while their own use of these tools is draining the information from written applications and unsupervised tests.
Most conversations about automated hiring start from the employer's desk and stay there. Three assumptions sit underneath them. The first is that candidates will accept automated screening once they see it is quick and consistent. The second is that a candidate who uses a chatbot is simply cheating, so the fix is a detector and a stern warning. The third is that a polished, well tailored application still tells you something real about the person who sent it.
Each assumption is partly wrong, and the evidence pulls them apart in different directions. Applicants judge a selection process by whether it feels fair and relevant to the job, and being assessed by software often fails that test, although the newest field evidence suggests the resistance is softer than laboratory studies imply. Candidate use of writing tools is not always cheating: in one large experiment, cleaner writing helped employers spot good workers. Yet once writing becomes free to produce, the effort that used to separate a careful applicant from a careless one disappears, and employers lose a source of information they relied on for decades. This article sets out what the research says on both sides of that arms race and what it means for the people who run hiring.
What Do Most Employers Get Wrong About AI in Recruiting?
The usual story treats automation as a one way street. The employer buys software, the software screens applicants, and the applicants either pass or fail. In that story the candidate is a passive input. Nothing in the research supports that picture.
Candidates are active. They form views about whether a process is fair, and those views shape whether they accept offers and whether they speak well of the employer afterwards. They also adapt. When a hiring process rewards a particular kind of written answer, candidates find ways to produce that answer more cheaply. Large language models, which are programs trained on huge amounts of text to predict the next words in a sentence and which power the popular chatbots, have made that adaptation almost free.
The second mistake is to frame candidate tool use only as a question of honesty. Some of it is dishonest, such as having a chatbot sit a personality questionnaire on your behalf. A great deal of it is closer to using a spell checker or asking a friend to read your cover letter. The moral framing hides the more important problem, which is about information. Hiring works because some things are hard to fake. A tailored cover letter used to take time and care, so only motivated applicants wrote good ones. When a machine writes a tailored letter in seconds, the letter still looks good but it no longer tells you who cared.
The third mistake is to assume that what holds in a laboratory holds in a real hiring process. Much of the evidence on how applicants react to automated selection comes from people imagining a scenario or watching a video. Real applicants with real jobs at stake may behave differently, and the evidence below shows that they sometimes do.
How Have Candidates Always Judged the Way They Are Selected?
Before asking how applicants react to software, it helps to know how they react to selection methods in general. Researchers have studied this for decades, and the findings are stable.
One influential meta analysis, which is a study that pools the results of many earlier studies, combined 86 samples covering 48,750 people. Applicants who saw a selection process as fair were more likely to think well of the organization, to say they would accept an offer and to recommend the employer to others. The strongest driver of those fairness judgments was job relatedness: whether the method looked connected to the work and seemed likely to predict performance. Interviews and work samples, in which candidates perform a slice of the actual job, were rated most favorably. Personality questionnaires and biographical questionnaires sat lower, and handwriting analysis came last.
That same study carried a warning that matters for everything that follows. The link between fairness perceptions and what people said they would do was stronger in hypothetical settings than in real ones. In other words, people imagining a selection process report stronger reactions than people actually going through one.
A later meta analysis pulled together 38 samples from 17 countries and compared reactions to 10 popular assessment methods. The ranking held across countries. Work samples and interviews were the most preferred, resumes, ability tests, references and personality questionnaires formed a middle group, and honesty tests, personal contacts and handwriting analysis were least liked. The authors also found that methods applicants liked tended to be the methods with better evidence of predicting job performance.
Two lessons from this older work carry straight into the present. Candidates want to be judged by something that looks like the job, and they want a human conversation somewhere in the process. Automated assessment threatens both at once, which is why the next body of research matters.
Do Applicants Trust a Machine to Interview and Judge Them?
The short answer from controlled studies is no, or at least less than they trust a person.
One early experiment asked 123 participants to watch and rate videos of either a highly automated interview, in which software conducted and evaluated the interview, or an interview by videoconference with a human. The researchers also varied the stakes, presenting the interview either as a hiring decision or as a training exercise. People accepted the automated interview less. They felt less social presence, meaning less sense that another person was there with them, and they saw the process as less fair. When the stakes were high, the automated version also left people feeling more uncertain and less in control.
A qualitative interview study went deeper with 33 professionals who were asked about artificial intelligence evaluating job interviews. They raised every kind of fairness concern researchers usually measure: whether the outcome was fair, whether the procedure was fair, whether they would be treated with respect and whether they would be told enough. Many accepted that software might be more objective than a person. They still wanted a human involved. The authors described this as preferring the human bias you know over an algorithm you do not understand. Participants also read the choice of tool as a message about the employer, so automated evaluation sent a signal about the organization before any decision was made.
Four online studies run with paid participants tested this more formally. People rated hiring processes driven by algorithms as less fair than processes run only by humans or by humans assisted by algorithms. This held whether the outcome was good or bad for the applicant. The proposed reason was that people believe an algorithm cannot recognize what makes them unique as a candidate.
Two further studies with 172 and 276 participants tested reactions to the scoring of a recorded video interview and a personality questionnaire. Simply telling participants that an algorithm, rather than a recruiter, had scored their answers made the process feel creepier, lowered how fair and accurate they thought it was and made them less willing to accept the feedback. It also lowered their attraction to the organization and their stated willingness to accept a job offer. Getting a favorable decision did not cancel out these effects. Candidates who passed still liked the algorithm less.
The most complete picture comes from a 2026 meta analysis of 53 studies covering 24,578 participants. Decisions made by algorithms in human resources settings were linked with less favorable reactions toward the system itself, including lower trust and perceived fairness, and toward the organization, including lower attractiveness and weaker intentions to pursue a job there. The size of the penalty depended on how people interacted with the system and on how much of the decision the algorithm made. That last point is the useful one. A tool that supports a human decision draws less resistance than a tool that makes the decision.
So the controlled evidence is consistent. Applicants dislike being judged by software, they dislike it more when the stakes are high and the machine has the final word, and the dislike spills over onto the employer.
Is the Resistance to Automated Interviews Softening?
Here the evidence becomes mixed, and it is worth being careful about what each kind of study can and cannot show.
The newest and largest piece of evidence comes from a natural field experiment, meaning a randomized test inside a real hiring operation where applicants did not know they were in a study. It involved about 70,000 applications for entry level customer service jobs handled by a recruitment firm in the Philippines in 2025. Applicants were randomly assigned to a first interview with either a human recruiter or a voice agent, a computer program that held a spoken conversation. Human recruiters then reviewed the interviews and made the hiring decisions. Applicants interviewed by the voice agent were 12% more likely to receive an offer, more of them started work and they stayed longer, with no drop in productivity. The transcripts suggested why: the automated interviews were more structured and consistent while still responding to what each applicant said.
The candidate side of that study is the surprise. When one group of applicants was allowed to choose their interviewer, 78% chose the voice agent. In a follow up survey, applicants rated their stress, comfort and the quality of feedback about the same in both conditions, and they accepted offers at similar rates. The one clear complaint was that the automated interview felt less natural.
This is a single study, it has not yet completed peer review and it comes from one country, one job type and one firm. Only 14% of invited applicants answered the survey, so the experience ratings may not represent everyone. The study also found that applicants who chose the voice agent tended to score lower on language and analytical tests than those who chose a human, which suggests that weaker candidates may prefer the machine. Even so, it fits the warning in the older meta analysis. People who imagine an automated interview react more strongly than people who sit one for a real job they want.
How do these findings fit together? The most honest reading is that applicants care less about whether a machine is involved and more about how the process treats them. The laboratory studies tested fully automated judgment presented as a black box. The field experiment used automation for a structured first conversation, with humans making the decision. That is the exact boundary condition the 2026 meta analysis identified. Handing the decision to software draws resistance; using software to gather information consistently for a human to judge appears to draw much less.
What Happens When Candidates Write Their Applications With a Chatbot?
The other half of the arms race starts on the candidate's laptop. Applicants now use writing tools to produce resumes, cover letters and answers to screening questions, and the research on what this does is more nuanced than the panic suggests.
A large field experiment on an online labor market randomly gave nearly half a million new jobseekers access to algorithmic writing help on their resumes. This was an older style of tool that corrected errors and improved readability rather than writing text from scratch. Jobseekers who got the help were hired 8% more often and at 10% higher wages. Employers who hired them were no less satisfied with their work. The authors concluded that writing quality is an imperfect sign of ability, and that clearer writing helped employers see the ability that was already there. They also disclosed that two of the authors received funding from the platform where the experiment ran.
That finding matters because it cuts against the reflex that any machine help is fraud. When an applicant's poor spelling hides real skill, a tool that removes the noise helps everyone, including the employer.
The harder question is whether recruiters can even tell when a machine wrote the words. In six experiments, 4,600 participants were asked to judge whether short personal self descriptions, including professional profiles of the kind found in applications, had been written by a person or by a language model. People could not reliably tell. Worse, they relied on intuitive rules that do not work, such as treating first person pronouns, informal contractions or mentions of family as signs of a human writer. Because those rules are predictable, a language model can be prompted to sound more human than a human. A recruiter who believes they can spot machine writing by feel is, on this evidence, mostly guessing.
Is the Cover Letter Dead as a Signal?
To answer this, it helps to understand what economists mean by a signal. A signal is something a candidate does that costs effort, where the cost is lower for stronger candidates, so the act itself carries information. A cover letter that clearly speaks to a specific job used to be a decent signal. Writing it took time and thought, and the applicants willing and able to do it well tended to be the ones more likely to do the job well. Generative tools remove the cost. Two recent studies from large freelance work platforms show what happens next.
The first working paper, which has not yet completed peer review, studied a platform that launched a built in tool in April 2023 that drafted a tailored cover letter in one click. The data covered more than 5 million bids from over 260,000 freelancers. Among workers who used the tool, the chance of getting a callback rose by about 3.6 percentage points from a starting rate of about 7%, roughly half again as many callbacks. That boost was statistically weak by the usual standards and faded after about two months. The bigger finding concerned employers. After the launch, the link between how well a cover letter matched the job and whether the applicant got a callback fell by 51%, and the link with receiving an offer fell by 79%. Employers turned instead to harder to fake information, especially the worker's past record and reviews on the platform. Workers who spent more time editing the machine draft did better, which suggests effort still counted where it remained visible.
The second working paper, also not yet peer reviewed, followed about 61,000 coding job posts and 2.7 million applications on a freelance platform from 2021 to 2024. Before chatbots, employers placed real value on tailored applications: a clearly more customized proposal was worth about as much to them as a bid around $26 cheaper. After chatbots arrived, that premium largely vanished, and tailoring no longer predicted whether the hired worker would finish the job. Applications per job more than doubled, from about 29 to 63, as writing them became cheap. The authors then used an economic model to simulate a world where written applications carry no information about ability. In that world, workers in the top fifth of ability were hired 19% less often and workers in the bottom fifth were hired 14% more often. That simulation is a model projection, not an observed outcome, but it shows where the logic leads. When everyone's application looks excellent, the people who lose most are the genuinely excellent.
Put the two papers beside the resume experiment and a clear pattern emerges. Help that removes noise from a candidate's own writing can make applications more informative. Help that replaces the candidate's effort makes them less informative. The cover letter is not dead as a piece of paper, but as a signal of motivation and fit it is badly wounded, and recruiters who keep scoring it as before are measuring the chatbot rather than the person.
Can a Chatbot Fake a Personality Test or Pass an Ability Test?
Personality questionnaires used in hiring have always faced a faking problem. An applicant who wants a sales job can guess that a sales employer wants someone outgoing and organized, and answer accordingly. The question now is whether a chatbot can do this better than a motivated human.
An early test prompted four freely available language models in mid 2023 to give ideal answers on personality measures for a sales job, then compared them with 869 university students who had been told to fake. Two formats were used. In the first, people rate how much each statement describes them, which is easy to fake. In the second, called forced choice, people must pick between two equally attractive descriptions, which is harder to fake. The best model scored higher than 99.6% of the faking students on the easy format and higher than 91.8% on the harder one. Most of the other models struggled more with the forced choice format.
A 2025 follow up gave a more cautious result. It compared 100 working adults with 100 sets of chatbot answers and tested two detection methods: a scale that flags people who describe themselves as implausibly virtuous, and an overclaiming questionnaire, which catches people who claim to know about things that do not exist. Without coaching, the chatbot did not consistently outperform humans at faking. When the chatbot was coached, however, it became better at avoiding the overclaiming trap. The authors concluded that current detection tools have real limits.
Ability tests, which have right and wrong answers, are a different matter, and here the change has been fast. One 2025 commentary reported a striking update. On a number series test, a measure of quantitative reasoning, an earlier widely used language model scored below the 20th percentile against thousands of human test takers, meaning that more than 80% of people did better. A newer reasoning model released in late 2024 scored at the 95th percentile, beating 95% of people. The author argued that no unsupervised test is likely to stay safe for long and set out options: supervised testing, strict time limits, detection software, asking candidates to think aloud, analyzing how candidates move through a test, favoring work samples over abstract tests and redesigning tests so that candidates are allowed to use the tools.
Not everyone agrees about how much damage is already done. A 2025 review borrowed from older research on test coaching and sketched three scenarios for what applicant chatbot use could do to average scores and to how well scores predict later job performance. The authors argued that these scenarios are not equally worrying, that each calls for a different response, and that applicant use can carry benefits as well as costs.
The most reassuring evidence so far comes from real applicants. An archival study published in 2026 examined large samples of job applicants who took two ability tests and one personality questionnaire before and after the public launch of a widely used chatbot in November 2022. Average scores barely moved. The changes were close to zero and pointed in different directions across tests. The lead researcher works for a test publisher, which is worth knowing, and the absence of a shift in averages does not prove that no individual cheated. Still, it is the best available evidence against the claim that unsupervised testing has already collapsed.
The fair summary is this. Language models can now fake personality questionnaires and solve ability items well enough to beat most humans under laboratory conditions. Real applicant data has not yet shown widespread score inflation. The gap between what is possible and what is happening is unlikely to last, because the tools keep getting better and easier to use.
Can Employers Design Assessments That Resist the Chatbot?
If candidates can outsource writing and test taking, the obvious response is to measure things that are harder to outsource. Two studies show both the promise and the trap.
A conference experiment with 400 participants, published in peer reviewed information systems proceedings, asked people to describe themselves in writing, much like a cover letter, and also to complete a standard Big Five personality questionnaire, which measures five broad traits including how organized and how agreeable someone is. The researchers then watched how cooperatively participants actually behaved. When people had an incentive to present themselves well, models that read their free written descriptions predicted cooperative behavior better than models built on the questionnaire. A language model could also detect, from the text, when people had a reason to fake. The authors titled their paper around a revival of the cover letter. The irony is that this study ran before chatbots could write the self description for the candidate. The same free text that resists faking when a person writes it becomes easy to fake once a machine writes it.
A 2025 chatbot study in Serbia and Montenegro tested a conversational chatbot that asked 159 people open questions to assess their personality, including 45 real job candidates. The chatbot's scores were less distorted by flattering self presentation than a traditional questionnaire. That is the good news. The bad news is that the chatbot measured only two of the five traits reasonably well and its scores did not predict real outcomes such as education level or job position. Resisting faking is useless if the measure stops predicting anything.
The lesson is uncomfortable but clear. Making an assessment harder to fake often makes it harder to validate as well. The approaches most likely to hold up are the ones that applicants already rate as fair: supervised work samples and structured conversations where the candidate must think and respond in real time.
What Does AI in Recruiting Mean for You as a Recruiter?
If you run hiring, the research points to a shift in where you look for information about candidates.
Start with your written screens. If you still score cover letters or open text answers for tailoring and polish, you are now measuring access to a chatbot and the skill of prompting it. You do not need to ban the tools to fix this. You need to stop treating written polish as evidence of motivation and fit, and move that weight to things that remain hard to fake, such as verified work history, references and a short task done under observation.
Next, look at your unsupervised tests. The archival evidence says scores have not yet inflated across the board, so there is no need to panic and scrap every online assessment. But the laboratory evidence says the tools can already beat most humans, so you should plan for supervised or time limited retesting of shortlisted candidates, at least for roles where the test carries real weight in the decision.
Then think about how candidates experience your process. Applicants consistently dislike being judged entirely by software, and they hold it against your organization even when they pass. They react far less badly when automation gathers information in a consistent way and a person makes the call. If you use an automated first interview, tell candidates that a human reviews every recording and makes the decision, and make sure that is true.
Finally, do not trust your instincts about machine written text. People are poor detectors, and the cues they rely on are the very cues a language model can mimic. Build your process so that it does not depend on catching anyone.
Key Takeaways
- Candidates react to the design of a hiring process, not only its outcome, and a process where software makes the final call draws more distrust and harms the employer's attractiveness, even among candidates who are offered the job.
- Fresh field evidence suggests that applicants accept automated first interviews far more readily when real jobs are at stake and humans still decide, so laboratory findings probably overstate the backlash.
- Writing help that cleans up a candidate's own words can make applications more useful to employers, while tools that write the content for them strip out the effort that once made tailored applications informative.
- Tailored cover letters lost much of their predictive value on large freelance platforms once one click writing tools arrived, and employers there switched to work records and reviews instead.
- Language models can fake personality questionnaires and solve ability items well enough to beat most people in experiments, yet large samples of real applicants have so far shown little average score inflation.
- Neither recruiters nor simple rules of thumb can reliably tell machine written text from human writing, so a hiring process built around catching cheats is built on sand.
Implications for Practice
Reweight your written screening. The freelance platform studies show that tailoring and polish stopped predicting callbacks and completed work once writing tools became common. Keep cover letters if candidates expect them, but give them little or no weight in shortlisting. Shift that weight toward information that is costly to fake: verified employment history, concrete examples you can check with references and short work samples done live.
Treat unsupervised tests as a first filter rather than a final verdict. The ability test evidence shows that newer reasoning models can now outscore 95% of people on some item types, even though real applicant averages have not yet moved. For roles where a test score carries real weight, add a short supervised or strictly timed check for shortlisted candidates, and compare the two scores. Large gaps deserve a conversation, not an accusation.
Choose forced choice personality formats over simple rating scales where you use personality measures for selection, since chatbots found them harder to fake, and combine any faking detection scale with other evidence rather than rejecting candidates on the scale alone. The 2025 faking study showed that a coached chatbot can learn to dodge at least one detection method.
Keep a human visibly in charge of decisions. The meta analytic evidence shows that resistance grows as the algorithm takes more of the decision, and the field experiment shows that applicants accept automated information gathering when people decide. Say so plainly in your candidate communications, give candidates the option of a human interviewer where you can, and then honor whatever choice they make.
Write a clear policy for candidate tool use and publish it with the job advertisement. Tell applicants which stages allow writing help, such as tidying a resume, and which do not, such as a supervised reasoning test. The resume experiment suggests that permitting help with clarity costs you nothing and may help you see ability more clearly, while a vague policy leaves honest candidates at a disadvantage against those willing to bend the rules.
Finally, review your hiring data every six months. Track whether written screening scores still relate to interview performance and later job performance in your own organization. Published studies come from freelance platforms and controlled experiments, and your applicant pool may behave differently. Your own data is the only evidence that will tell you when a screening step has stopped working.
