Our view: using AI to write a résumé should not, by itself, decide whether someone gets an interview. A résumé makes claims about experience. Your hiring process needs a fair way to examine the claims that matter for the job, regardless of who helped write the document.
Which claim would change your hiring decision if it turned out to be unsupported?
01 / Separate presentation from experience.
“Led a migration” can describe very different responsibilities. One person may have chosen the architecture; another may have tested a single service. Both can have useful experience. A fluent summary does not tell you which contribution the candidate made.
Choose a responsibility in the role and ask for a concrete account: the starting problem, the person’s own work, a decision they made, and how they checked the result. Ask the same core questions of candidates being considered for the same responsibility. Leave room to explain unfamiliar terminology or a different route into the work.
02 / A text detector cannot establish someone’s competence.
NIST’s 2024 GenAI pilot examined generated summaries and text detection, finding performance varied across systems. That study did not validate a résumé-screening test. Its results should not be presented as proof that a detector can identify an unqualified applicant.
Our recommendation is to assess the experience itself. If AI use matters to the role, discuss it directly: what did the person use it for, what did they check, and what decisions remained theirs? An authorship score does not answer those questions. A claim that cannot yet be checked should remain an open question, not become an invented negative finding.
NIST: text generation and detection evaluation (opens in a new tab)03 / Give reviewers evidence they can examine.
For each important claim, retain the source, the candidate’s explanation and the reviewer’s reasoning. A small, role-relevant exercise can add useful evidence when it has clear instructions and consistent assessment criteria. Avoid asking candidates to produce unpaid work for a live customer.
Distinguish what you observed from what you inferred. “Explained how they rolled back a release” is an observation. “Will own every production incident independently” is a much broader inference. Record the remaining question before the hiring team meets.
- Responsibility being assessed
- Claim and source
- Question or work sample
- Observed evidence
- Uncertainty and next question
04 / Make the hiring decision understandable.
An AI-assisted process can organize applications and prepare review material. The hiring team still needs to understand why someone advances. Measure whether reviewers receive useful evidence and whether the assessment relates to the work. Counting processed résumés alone misses that purpose.
Start with one role and a shared evidence brief. Compare how reviewers interpret it, resolve ambiguous criteria, and keep a route for candidates to clarify missing information. Improve the assessment before increasing the volume.
Review a résumé claim against the work
- Responsibility
- Maintain a customer-facing API.
- Claim
- Led a service migration.
- Question
- Which part did you own, what changed, and how did you verify it?
- Evidence to examine
- A clear account of a decision, a test and a response to an unexpected result.
- Unresolved
- Whether the person owned the release or contributed to one part; clarify before drawing a conclusion.