How widespread it is
Two recent surveys give the picture, both self-reported, so treat them with some caution:
- In a survey of 500 job seekers, 36% used AI in their job search; in the youngest age group, 46.7%. 54.2% would never mention it to an employer, 12.8% always say so. One in five lets AI submit applications automatically, and one in ten has been fed an AI prompt during a live interview.
- From the other side: according to a survey of 2,950 candidates, 63% have already been through an AI interview — 13 percentage points more in half a year. 70% were not clearly told in advance that AI would assess them; 21% only worked it out at the start of the interview. Just 18% say their employer has a transparent AI policy, and 57% think disclosure should be made mandatory by law.
So both sides use it, and neither side says so.
The detectors: between 33 and 81%
This is the most important measurement in the article, because most of the fear and most of the confidence start from this one point.
An academic study tested 14 detection tools on 54 documents across six categories, among them: human text, machine-translated human text, AI-generated, manually edited AI and machine-paraphrased AI.
| What was measured | Result |
|---|---|
| Overall accuracy | 33–81% |
| The best tool | 76–81% |
| False accusation on human text | 0–50%, 0% at seven of the tools |
| False accusation on machine-translated human text | 11.1% on average |
| Accuracy on manually edited AI text | ~30% |
| Accuracy on machine-paraphrased AI text | ~15% |
The authors' conclusion leaves little room for doubt: the detection tools get it wrong, and are neither accurate nor reliable — not even the best tool reached the eighty per cent the researchers treat as acceptable on every mode of evaluation.
Two things follow from this, pointing in opposite directions.
The first: AI-written text is not really being caught — after manual editing the hit rate sits around thirty per cent, after paraphrasing around fifteen.
The second one concerns anyone who does not write in English as a first language: on plain human text seven of the tools accused nobody at all, while on machine-translated human text false accusations averaged 11.1%. If you draft in your own language and then translate into English, you land in exactly the category where detectors accuse the innocent most often on average. Your own honestly written material can fail on a bad tool.
The price of owning up — and of being found out
This is the ethical spine of the article, and it helps that it is measurement rather than opinion.
13 experiments, more than 5,000 participants. The question was what happens to trust when somebody admits to having used AI.
| Situation | Change in trust |
|---|---|
| A teacher admits to grading with AI | −16% |
| A company discloses AI use in its advert | −18% |
| A designer tells the client | −20% |
The effect showed up in every single experiment, including where the evaluators were themselves frequent AI users.
And the bad news: according to the researchers, trust falls even further when somebody else exposes it — the worst possible outcome.
So the position is uncomfortable: owning up costs you, being found out costs you more, and the detector is unreliable — and the employer will use it regardless.
The conclusion is not that you should hide. It is that the content should genuinely be yours: let AI help with the phrasing, not with the facts. If you could not say it out loud, do not put it in.
What is actually banned
Concrete employer policies are hard to find — but there is one well-documented and instructive reversal.
A large technology company banned AI in applications until May 2025. The reasoning was that this was how it could judge a candidate's personal interest and their communication skills unassisted by AI.
In July 2025 the ban was withdrawn. Polishing — CV, cover letter, application materials — has been permitted ever since. It is still forbidden in most assessment exercises and in the live interview, unless it is explicitly allowed; for preparation (research, practice) it is free.
The reasoning behind the turn was practical: the ban proved hard to enforce, the company itself uses AI in its hiring process, and giving candidates access levels the field.
The two positions together show today's situation well: a ban cannot be enforced, so the rules moved to where they can be checked — to the live interview and to the assessment exercise.
Does it improve your odds?
There is one measurement with a very large sample, and a very important piece of small print attached to it.
In a field experiment on an online labour market with nearly half a million job seekers, the treated group got algorithmic writing help for their CV. The probability of being hired rose by 8%, and the authors explicitly point out that they found no sign of employers being less satisfied.
But this was spelling and grammar correction, not a CV rewritten by AI. So the +8% proves that bad writing measurably hurts and fixing it measurably helps — not that you should have the machine write your CV. The two must not be blurred together.
There is even a signal pointing the other way. A study running across 880,000 texts found that language models cut the variance of written complexity by 21–50% — reinforcing conformity at the expense of distinctiveness. If everybody has their CV written by the same tool, those CVs start to resemble each other, and what is lost is precisely the power to stand out. (This has not been measured on CVs directly — it is a transferred conclusion.)
The legal side
This part is in motion in 2026, so it is worth being precise.
Recruitment is a high-risk use. The annex to the EU AI Act explicitly places here the systems that post targeted job adverts, filter applications and evaluate candidates, as well as those that decide on promotion, termination of employment or the allocation of tasks, or that monitor performance and behaviour.
But the deadline slipped. The obligations covering standalone high-risk systems were originally to apply from 2 August 2026; the amendment that took effect in the summer of 2026 pushed them out to 2 December 2027.
Two things are already mandatory, though:
- Transparency. Under Article 50 of the Act, anyone whose AI system enters into direct contact with a person has to inform them clearly and distinguishably, at the latest at the first interaction. This has applied since 2 August 2026, and the postponement did not touch it. So if an AI interviewer is talking to the candidate, that has to be said.
- AI literacy. Under Article 4 of the Act, since 2 February 2025 the organisations operating these systems have had to take measures to build the AI literacy of the staff involved in the process.
On top of that, the data protection regulation's rule on automated decision-making still applies unchanged, independently of the AI Act.
In Hungary the implementing framework is set by Act LXXV of 2025 and Government Decree 344/2025 (X. 31.), and market surveillance is carried out by the AI Market Surveillance Authority, which is not a standalone body but an organisational unit inside a ministry. I found no Hungarian official position or case specifically about AI in recruitment — which does not mean there is none, only that it is not publicly available.
(This is not legal advice. A specific situation is decided by the classification, and that is worth going through with a professional.)
The line: what counts as help and what counts as cheating
A surprisingly simple rule comes out of the measurements above.
HELP
- Grammar, phrasing, structure
- Finding the keywords in the job advert
- Condensing your own experience
- Preparation: questions, practice, company research
- Translation - but from your own sentences
CHEATING
- Experience that is not there
- A result you did not achieve
- An answer fed to you in a live interview
- An assessment task you did not solve
- An auto-submitted application you never read
The dividing line is not the tool, it is the claim. If the text claims something about you that is not true, it does not get any better for the fact that you could have typed something similar yourself.
A practical self-test: read your application out loud and at every sentence ask whether you could say this live too, with an example. Cross out whatever you cannot. It is the same sentence that will come up in the interview anyway.
Common misconceptions
- "The detector will filter it out." It will not. Even the best tool only hovers around eighty per cent, and on manually edited text it is around 30%.
- "They can't accuse me, I wrote it myself." They can. On machine-translated human text false accusations averaged 11.1%, while on plain human text seven of the tools never accused anyone at all — which is exactly the case of the applicant who is not a native English speaker.
- "Better to own up." Trust falls anyway by the measurements, by 16–20%. But being found out is worse. So the strategy is not the disclosure; it is that the content should be true.
- "They measured that a CV written with AI does better." What they measured was spelling and grammar correction, not a rewrite. The +8% refers to that.
- "Nothing applies to recruitment AI yet." The transparency obligation has been live since August 2026, the AI literacy one since February 2025. Only the full high-risk package slipped to December 2027.
What to take away
With an application the question is not whether you use AI — by the measurements the other side uses it too, and the detectors will not settle it anyway.
The question is what you claim. Phrasing, structure, grammar and preparation you can hand over without worry: those are exactly the areas with the best evidence behind them. Experience, results and live answers you cannot — not for ethical reasons, but because the interview will bring it out regardless.
And one thing worth keeping in mind: if everybody has the same tool write for them, sameness itself becomes the disadvantage. Your own voice is worth more today than ever.
Which part of the writing is worth handing to a machine is measured out in the faster reading post, study by study; and building the daily habit is the subject of the weekly AI routine post, with a schedule.