Using AI on a job application is normal and mostly fine. The trouble is that both sides are now using it, and one side is using it badly — often by trusting a detector.
- Writing help is not cheating. Better sentences describing real experience is what an editor does.
- The line is the facts. Adding a skill, a year, or a result you did not produce is lying on a document you will be asked about.
- The interview is the audit. Anything on the page can be asked about, and generated achievements do not survive a follow-up question.
- Employers screening with AI detectors will misjudge people — the evidence is strongest against non-native English speakers.
- Generic is the real failure mode, not detection. An obviously templated letter loses on being boring long before anyone runs a checker.
- Keep your own voice in it. The parts a model smooths away are usually the parts that made you distinctive.
A model can help you say what you did. It cannot know what you did.
Everything that goes wrong on an AI-assisted application comes from letting it write the second half.
Where the line actually is
The distinction is not about effort or authenticity. It is about whether the document is true. A CV is a factual representation, and in many places a materially false one is grounds for withdrawing an offer or dismissal later — years later, when it surfaces.
Ask one question of every line: could I talk about this for two minutes if asked? If not, it should not be on the page — regardless of who wrote the sentence.
The problem on the employer's side
Detectors misjudge people, and they misjudge some people more
Employers increasingly run applications through AI detectors. Those tools are not reliable enough to support a decision about an individual, and their errors are not evenly spread: peer-reviewed testing found detectors wrongly flag more than half of writing by non-native English speakers as machine-generated — an average false-positive rate of 61.3% across seven detectors on 91 TOEFL essays, against near-perfect accuracy on essays by US eighth-graders. And 97.8% of the TOEFL essays were flagged by at least one detector.Corrected 11 Sep 2026: this sentence said detectors misclassify "a substantial majority" of non-native writing. The study says more than half. Liang, Yuksekgonul, Mao, Wu & Zou, "GPT detectors are biased against non-native English writers", Patterns 4(7):100779, July 2023 — cell.com returned 403 to this check, so not read at source there. The study itself was read at source 17 Sep 2026 in its full text on Europe PMC: “In our study, we evaluated the performance of seven widely used GPT detectors on 91 TOEFL (Test of English as a Foreign Language) essays from a Chinese forum and 88 US eighth-grade essays from the Hewlett Foundation’s ASAP dataset.” “While the detectors accurately classified the US student essays, they incorrectly labeled more than half of the TOEFL essays as "AI-generated" (average false-positive rate: 61.3%).” “All detectors unanimously identified 19.8% of the human-written TOEFL essays as AI authored, and at least one detector flagged 97.8% of TOEFL essays as AI generated.” Its Figure 1 caption reports “near-perfect accuracy for US eighth-grade essays”. Also read at source 11 Sep 2026 in Stanford's own write-up, ScienceDaily, 10 Jul 2023: "These platforms incorrectly labeled more than half of the essays as AI-generated, with one detector flagging nearly 98% of these essays as written by AI," while "The detectors were able to correctly classify more than 90% of essays written by eighth-grade students from the U.S. as human-generated." ScienceDaily’s “one detector flagging nearly 98%” is not what the paper says: 97.8% is the share flagged by at least one of the seven.
The pattern is worth understanding because it explains the unfairness. Detectors key on features like limited vocabulary variety and even sentence structure — which is also what careful writing in a second language looks like.The mechanism, read at source 11 Sep 2026 in the same ScienceDaily report, quoting the study's authors: "If you use common English words, the detectors will give a low perplexity score, meaning my essay is likely to be flagged as AI-generated. If you use complex and fancier words, then it's more likely to be classified as human written by the algorithms." Further evidence on detector false-positive rates, with dates, on how to spot AI writing. Nothing here claims a specific employer uses one; the point is that the tools cannot support an individual judgement. Checked 6 Sep 2026, re-checked 11 Sep 2026
You cannot control this, which is frustrating and worth saying plainly. What you can do is give a reader reasons to believe a person wrote it: specifics only you would know, an unusual detail, a sentence that is yours.
If you are on the hiring side
Do not screen on a detector score. It will reject candidates for writing English as a second language, which is both unfair and the kind of thing that becomes a legal problem. Screen on the work and on the interview, which is where invented experience surfaces anyway and cannot be faked in real time.
A team AI policy you can actually adopt covers the same argument for internal work: enforce on disclosure and quality, never on a percentage.
Generic loses before detection does
The realistic risk is not being caught. It is being forgettable. A letter that opens "I am writing to express my enthusiastic interest in this exciting opportunity" is not rejected for being AI — it is rejected for being the ninetieth of its kind that morning.
What survives a skim is specificity: the thing you actually did, the reason you want this job and not a similar one, something about them that could not be said about any other employer. A model does not know any of that, which is precisely why the letters it writes unaided are interchangeable.
Use it to fix your sentences, not to have your thoughts.
1. Write the ugly version yourself. Bullet points, no polish, everything true.
2. Ask for tightening only — shorter, clearer, active. Explicitly forbid adding facts.
3. Put back one thing it smoothed out. There is usually a specific, slightly odd detail that made you sound like a person.
4. Read it aloud. If it does not sound like you, it will not sound like you in the room either.
Step two is the instruction people skip — "do not add anything I did not say" — and it is what stops the invented metric appearing.
Before you send it
Detector reliability, including the false-positive finding for non-native English writers, is sourced with dates on how to spot AI writing. The underlying study is Liang, Yuksekgonul, Mao, Wu & Zou, "GPT detectors are biased against non-native English writers", Patterns 4(7):100779, July 2023 — cell.com blocks automated fetches and was not read at source here; the figures above were read at source in the study’s full text on Europe PMC on 17 Sep 2026, and the summary in Stanford's write-up via ScienceDaily on 11 Sep 2026. Checked 6 September 2026, re-checked 11 and 17 September 2026.
No employment law is quoted here and no jurisdiction is assumed. How a false statement on an application is treated varies by country and by contract; the general position that it can justify withdrawing an offer or later dismissal is widely held rather than sourced to one statute. This is not legal or careers advice — the durable part is that an interview tests everything the document claims, and that is true everywhere.
The through-line: let it improve how you say things and never what you say. The interview is coming, and it only asks about the second one.