Is it safe to let AI rewrite your resume? Here's what it makes up
Not without a fact-check layer, and the thing it most reliably makes up is experience years. Language models don't know today's date and don't do date arithmetic — they pattern-match. From our own production logs: a model counted a 2016–2020 bachelor's degree as four years of work experience; a fabrication filter turned the award year "2018" into "1 year of experience"; and a rewrite produced "5 years of product management" for a candidate with zero PM roles, because her total experience happened to compute to 5.3 years and the number 5 looked authorized.
Why it happens
A model asked to "strengthen" a résumé optimizes for what strong résumés look like, and strong résumés contain confident numbers. It cannot verify them: it doesn't know the current date, so "Nov 2020 – present" is not computable; it doesn't distinguish a calendar year from a duration, so "2018" and "1 year" share a unit; and it evades naive filters effortlessly — ban digits and it writes "nearly a decade", which is the same fabrication in formal wear. We had to normalize spelled-out numbers back to digits before checking, and to explicitly exclude four-digit years from duration parsing, and each of those rules exists because a real résumé went wrong without it.
Prompt rules don't hold; gates do
"Never invent numbers" in the prompt fails probabilistically — it holds for twenty generations and breaks on the twenty-first. The reliable fix is a deterministic gate outside the model, at the single exit point every generated résumé passes through, that re-verifies every quantified claim unconditionally. Unconditionally matters: we first checked only claims we couldn't trace to a source, and a wrong-but-coincidentally-plausible "5 years" sailed through because the number 5 could be traced. Same-value coincidence is not verification.
The gate needs a third action: ask
A gate that can only edit or delete will decide things on your behalf — either inventing a corrected number (fabrication again, ours this time) or deleting a true sentence to punish a wrong number. The correct third action: strip the unverifiable quantity, keep the claim ("5 years of product management" → "Product management"), and put a question to the human with the evidence attached — here are your date ranges, which figure do you want to state? The material for the judgment goes to the person; the judgment stays with the person.
How to protect yourself with any AI tool
- Diff every AI draft against your real history line by line — treat the draft as a proposal you're signing, because you are.
- Check every number, including spelled-out ones ("nearly a decade", "dozens of") and every date range's classification (work vs education vs awards).
- Be suspicious of new quantities that appeared without you providing them — that's the signature failure.
- Prefer tools that show you what they changed and flag what they couldn't verify, over tools that hand back a clean final document.
OfferCoach runs the gate described above — normalize, verify unconditionally, strip-and-ask — at the one exit every document passes through. We publish the failure cases because they generalize: whatever tool you use, the questions in the list above are the audit.