With AI-assisted job applications, a polished document does not reliably tell you how much of the wording, structure or reasoning came from the candidate. That uncertainty is real. Treating tone as evidence of dishonesty creates a second problem: you may reward familiar writing styles, penalise legitimate editing or translation help, and still miss a fabricated story written by a person.
The practical response is to make application prose carry less weight. Decide what AI help is acceptable, ask for narrower evidence, probe one important claim and observe capability under conditions that resemble the actual role.
What you'll learn
- How to separate legitimate AI assistance from misrepresentation.
- What to state about AI use before candidates apply.
- How to turn a polished claim into evidence you can test.
- How to design a fair work sample for a role where AI may be allowed.
- Which job-related reasons can support a no-hire decision without guessing authorship.
What do AI-assisted job applications change?
That comparison does not prove AI caused the gap. It does show why more application volume cannot be treated as more hiring signal. In the same UK research, 27% of organisations that attempted to recruit reported what they considered excessive candidate use of generative AI. A hiring process built around generic written pitches now asks the document to prove more than it can.
This sharpens the distinction between confidence and competence in hiring. The old risk was mistaking a polished performance for ability. AI makes the polish cheaper, so the evidence behind it matters more.
Set the candidate AI rule before you assess anyone
A hidden rule invites inconsistent judgement. One interviewer may see a writing assistant as normal; another may treat the same use as dishonesty. The candidate cannot follow a boundary you never stated, and the hiring team cannot apply it consistently if everyone is carrying a private definition.
| Stage | Usually reasonable | Needs an explicit rule | Evidence to protect |
|---|---|---|---|
| Application form | Spelling, grammar, translation and structure help when the underlying claims remain accurate. | Whether generated first drafts or rewritten answers must be disclosed. | The candidate's real experience, contribution and results. |
| Role questions | Preparation and note organisation before submission. | Whether the final answer must be composed without live assistance. | Job-relevant judgement, not writing ornament. |
| Work sample | The tools people would normally use in the role, when tool use is part of the job. | Any restricted segment that tests an essential unaided skill. | Choices, verification, trade-offs and accountability for the result. |
| Interview | Reasonable preparation and accessible notes. | Whether live AI assistance is permitted during a specific exercise. | The person's ability to explain and adapt their own evidence. |
Use the application to confirm baseline eligibility and select decision-critical claims worth testing. Then probe one important claim, observe one short role-relevant task and score both against criteria written before the interview. For a fuller interview structure, see hire for values without culture-fit bias.
How do you test a polished candidate claim?
| Evidence step | Question | What a useful answer contains |
|---|---|---|
| Claim | You wrote that you reduced onboarding time by 30%. What changed? | A clear result with a defined unit and time period. |
| Context | What was happening before, and what constraints did the team face? | Baseline, team shape, available data and material limits. |
| Contribution | Which part did you own, and which parts belonged to other people? | Personal action without absorbing the whole team's work. |
| Choice | What alternative did you reject, and why? | A real trade-off, not a rehearsed sequence of best practices. |
| Outcome | How was the result measured, and what did not improve? | Measurement method, residual problem and proportionate confidence. |
| Transfer | What would you inspect first in our onboarding problem? | Relevant judgement adapted to this role rather than copied advice. |
A credible answer does not need confidential client files or a perfect memory for every number. It should hold together across sequence, ownership and trade-offs. If an important detail is uncertain, a strong candidate can name the uncertainty instead of manufacturing precision.
Design a work sample around the role's real tool conditions
A take-home assignment is not automatically better evidence. If it rewards unlimited unpaid time, polished decks or hidden outside help, it may reproduce the same uncertainty as the application. Keep it short, job-related and explicit about the tools allowed.
| Job condition | Assessment design | Score this |
|---|---|---|
| The role normally uses AI | Allow the relevant tool. Ask the candidate to show sources, checks, edits and what they would not delegate. | Judgement, verification, quality bar and accountability for the final work. |
| An unaided skill is essential | Run a narrow supervised segment and explain the restriction in advance. | Only the essential skill the restriction is designed to reveal. |
| Tools are mixed or uncertain | Use one allowed-tool task and a short follow-up where the candidate diagnoses a flaw or changes direction. | Whether capability survives a changed constraint. |
| The candidate needs an accommodation | Provide an equivalent route that preserves the capability being tested. | The same job criterion, not speed or format introduced by the assessment itself. |
When AI is allowed, review the candidate's output the same way a manager should review AI-assisted work before it becomes somebody else's review burden. Ask what was verified, which judgement stayed human and who owns the consequence if the answer is wrong.
Why should employers avoid AI detector scores?
OpenAI retired one of its own text classifiers because of low accuracy. In its published evaluation, that classifier identified 26% of AI-written challenge texts as likely AI-written and incorrectly labelled 9% of human texts. The page also warned that performance was worse on short text and outside English, and that editing could evade the classifier. Those figures describe one retired system, not every detector, but they show why a probability cannot carry a hiring decision by itself.
False-positive risk may not be evenly distributed. A 2023 study in Patterns found that the tested detectors disproportionately misclassified human essays by non-native English writers in its TOEFL dataset. Do not transplant that study's exact error rate to every resume or tool. Use it as a warning that writing style can become an unintended proxy for language background.
Even a detector that performs well in one benchmark cannot tell you whether the experience is true, whether the candidate made the key decision, or whether they can repeat the capability in your context. Those are the questions the hiring process exists to answer.
What evidence can support a no-hire decision?
- 01
Describe the concern without accusing intent
Point to the claim, contradiction, capability or material process rule that matters. Do not say the prose merely 'sounds like AI'.
- 02
Offer a fair clarification route
Ask the candidate to explain the work, their contribution and any tool use relevant to the stated boundary.
- 03
Apply the pre-written job criterion
Decide whether the evidence meets the role requirement or material process rule. Keep style and suspicion out of the score.
- 04
Record the evidence and uncertainty
Document the criterion, relevant evidence, clarification and remaining uncertainty in the same format used for other candidates.
A no-hire decision may still be justified for material misrepresentation or breach of a disclosed, job-relevant rule. Record the criterion and evidence without claiming authorship or dishonesty you cannot establish. 'The candidate could not substantiate the claimed result after clarification' is job-related. 'The application felt machine-written' is not. Check the hiring, privacy and discrimination rules that apply where you operate before turning process evidence into a decision.
After the hire, carry only job-relevant evidence into onboarding
Use those points to choose the first assignments and check-ins, then update them from observed work rather than treating the hiring record as a fixed judgement.
Make sure the role the person enters matches the one described during hiring; a strong candidate can still struggle when the lived role contradicts what they were hired to do.
Change one hiring stage before the next application arrives
Use this evidence prompt: 'Tell us about a difficult result relevant to this role and the part you personally owned.'
Follow with: 'What would you inspect first in this example problem from the role?' Put the AI-use boundary beside both questions so every candidate sees the same conditions.
You may never know exactly how every sentence was produced. You can still make a disciplined hiring decision. Reduce the value of polish, increase the value of verifiable work and keep the final judgement attached to the job.