A job applicant may describe their experience differently when they think a computer, rather than a person, will decide whether the CV deserves a second look. Research into AI CV screening suggests that, under experimental conditions, people can become more willing to exaggerate qualifications when they believe an automated system is evaluating them. The finding raises an awkward question for employers rushing to make recruitment faster: does the method of judging an application change the application itself?
The study does not prove that every applicant lies, nor that people using AI tools are inherently dishonest. It examines how candidates respond to the perceived evaluator. That difference matters because recruitment software is often assessed by whether it processes applications consistently or predicts job suitability. The possibility that applicants adapt their behaviour to the system complicates both goals. A fast filter is less useful if it changes the material being filtered.
Why AI CV screening might change what people write
When people apply for a job, they already choose what to emphasise. They might highlight relevant experience, omit unrelated work or describe a project in the most favourable accurate terms. The ethical line appears when that presentation turns into an unsupported claim about a qualification, responsibility or achievement. The researchers investigated whether knowing that the audience was an AI system, rather than a human assessor, affected how participants approached that line.
The academic work, described by its authors through an institutional research record, used a series of experiments to explore the effect. The reported pattern was that perceived AI evaluation could increase the inclination to embellish. The evidence concerns experimental responses and intentions under specified circumstances, not an audit of thousands of actual CVs submitted to employers. That is a meaningful result without being proof of a national epidemic of fraudulent applications.
One plausible reason is that applicants expect a machine to interpret claims differently. Someone may believe that an automated filter looks for keywords and pays less attention to nuance. If a candidate assumes that success depends on repeating the wording of an advertisement, they may be tempted to make their experience sound closer to the desired specification. The research does not justify treating this explanation as universally true, but it identifies a behavioural risk that employers cannot resolve simply by improving the classification algorithm.
The hiring problem employers already recognise
Most job descriptions combine essential requirements with qualities that would be helpful but not essential. A candidate who meets some but not all of them has to decide whether to apply and how to explain any gaps. That is normal. Trouble starts if the process rewards confident assertions over demonstrable capability. A screening system trained on historical CV data may amplify a familiar problem when it privileges certain phrases or credentials without enough context.
AI can also be used on the applicant’s side. Generative tools can reword a CV, suggest a clearer structure and help someone whose first language is not English present genuine achievements. None of that requires dishonesty. The problem is when assistance quietly becomes invention: adding a qualification that does not exist or attributing a team achievement to one person. Hiring managers need to distinguish good communication from unreliable content, regardless of who wrote the sentences.
LiveAIWire has explored privacy and fairness in AI recruitment and the way AI changes CV presentation and hiring decisions. The new research adds a behavioural dimension. A system can be statistically consistent in how it scores text and still face a changed pool of claims because people understand, or misunderstand, how to influence its decisions.
What the experiments establish and what they cannot
Researchers in behavioural studies can compare groups exposed to different descriptions of the same task. That helps isolate whether a perceived human or AI evaluator changes what participants say they would do. Such experiments are valuable because real applicants rarely volunteer to tell an employer when they exaggerated, and an employer may not know the truth even after checking a reference. But experimental responses are not the same as verified behaviour in the labour market.
The study is associated with a forthcoming journal issue, and its bibliographic details appear in a research publication record. The issue date is listed in 2027, so it would be misleading to describe the research as an event that happened that year. The relevant result is the study itself and its research findings, not a future calendar date. The public record should also be read for its study design, population and limitations before generalising to all occupations or countries.
There are several unanswered questions. Does the effect persist once applicants have actually experienced a company’s hiring system? Does it differ between graduates and experienced workers? Does it appear only when an application is read entirely by software, or also when a recruiter uses AI to rank candidates before a human review? Those distinctions affect what employers should do. A behavioural effect in an online experiment may be important without establishing its size in every real workplace.
Why keyword hunting is not the same as showing ability
An automated shortlist might depend on recognising specific skills, work history or qualifications. Applicants who understand the job may naturally use the correct terminology. A software developer with experience of a named programming language should say so. Yet a keyword match says little about depth. Somebody who used a tool once and somebody who built complex systems with it might both have the term in a CV. Screening that ignores context could invite applicants to maximise surface matches.
This is a design question as well as an ethical one. If an employer gives clear instructions, describes what the system checks and asks for verifiable examples, applicants may have less reason to guess what will please the filter. If the process is mysterious, people may rely on internet myths about secret keywords and machine-readable formats. A less confusing application process can benefit applicants and recruiters even without sophisticated AI.
The other danger is overreaction. Employers should not assume that a candidate who uses a writing assistant has something to hide. A candidate may use software to improve grammar, adapt a CV to a position or explain an employment gap more clearly. Penalising tool use itself could disadvantage people with language, literacy or accessibility needs. The meaningful standard is whether the underlying claims are accurate and relevant to the job.
The employer still has to verify the important claims
A well-designed selection process should test capabilities independently of CV wording. Work samples, structured interviews, practical demonstrations and appropriate reference checks can help, with care to avoid adding irrelevant or discriminatory barriers. No single method is flawless. The point is to avoid treating a polished document, written by a human or an AI tool, as proof that its contents are true.
Verification should be proportionate. An employer recruiting for a safety-critical role may need more rigorous checks than one seeking a short-term administrative assistant. Checks also have to comply with privacy and employment law. Companies should not collect broad amounts of personal information merely because an automated system can process it. Human oversight is useful only when reviewers have enough information and authority to question a doubtful outcome.
LiveAIWire previously examined how AI-assisted interviews can alter hiring speed and offers. Faster processing can reduce tedious waiting for candidates. However, if the software stage becomes an exercise in writing for a filter, apparent efficiency may hide extra work later when interviewers discover mismatches. Employers need to evaluate the entire route from application to successful employment, not simply the number of forms processed in an hour.
What applicants can reasonably take from this
For jobseekers, the safest and strongest strategy remains to describe real abilities clearly. Read the requirements, identify genuine matches and give concrete examples rather than padding the document with unsupported superlatives. If a tool suggests impressive language that exaggerates the role, edit it. Employers often follow up with specific questions, and a claim that cannot be explained comfortably may undermine otherwise strong experience.
It is sensible to ask how an application will be assessed, though companies may not disclose every internal detail. A candidate can still prepare a CV that a human can understand: straightforward headings, relevant examples and an explanation of why their experience suits the vacancy. These practices also tend to make information more legible to software. The goal should be accurate communication, not guessing a hidden scoring formula.
The research points to a feedback loop. An employer changes the evaluator; applicants change their presentation; the evaluator then receives different information. That is a problem no single accuracy percentage can capture. AI screening may still have useful roles in recruiting, particularly when it reduces routine administration. Its success will depend partly on whether employers design a process that rewards demonstrable ability rather than a performance staged for the machine.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.