AI·Apr 9, 2026·7 min read
Detecting AI-generated resumes: signals that actually work
Stylometric drift, employment-graph inconsistencies, and timestamp forensics — a practical guide for talent teams.
JN
Jordan Ng
ML Engineer, Trust & Safety
We've seen a 6x increase in AI-generated resumes year over year. Most are harmless polish — but a meaningful slice are entirely fabricated identities trying to land remote roles. Here's what we've learned about catching the bad ones without flagging the rest.
Signals that don't work
Forget perplexity scores and 'AI detector' models. They have ~60% accuracy on resume text and they punish non-native English speakers disproportionately. We don't use them and you shouldn't either.
Signals that do
- Employment graph inconsistencies — the candidate claims to have worked at X, but no one in their stated org cohort can be cross-referenced.
- Timestamp forensics — PDFs created seconds before submission, or font-rendering artifacts that match known LLM-to-PDF pipelines.
- Stylometric drift between sections — a resume where the summary, bullets, and cover letter were clearly written by different authors.
- Behavioral signals during the assessment — keystroke cadence, paste events, and tab-switch patterns.
How we surface it
Ethur's resume fraud module returns a structured signal set, not a single score. You see exactly which signals fired, with evidence, so your team can make a defensible decision.
The goal isn't to reject AI-assisted resumes. It's to catch fabricated ones — and the difference matters.