What our CV checker catches — and what it misses
We ran the free CV checker on 99 CVs whose authorship we knew. These are all the results, including the case where it fails.
Measured on 17 August 2026 with the prompt and model in production at the time (Claude Haiku 4.5).
Results
The checker scores each CV from 0 to 100 and flags it at 70 or above.
What it means
- It catches CVs pasted straight out of an AI: all 32 were flagged, against 1 of the 36 human CVs. A flag is worth acting on.
- It does not catch a candidate who asks the AI to sound human: none of the 31 were flagged. That takes one sentence and costs the candidate nothing.
- So a low score proves nothing. It means the signs of unedited AI text are missing, which is just as true of a careful person as of a careful prompt.
How the test was built
Each AI CV was written from a factual summary of one human CV — role, seniority, employers, skills — never from the human text itself, so the model could not copy its style. Paired this way, the three groups cover the same jobs, and the checker cannot tell them apart by subject matter.
Limits of this test
- The human CVs are from before 2022, so older conventions may read as human for reasons unrelated to who wrote them. The clean test is a recent real pile, which needs a customer's agreement.
- One instruction was used for the 'sound human' group. Other wording may score differently.
- All AI CVs came from one model family. Another model may score differently.
- 99 CVs is enough to show the direction clearly, not to quote a rate to the exact percent.
Why our assessments don't judge style
Style can be faked with one sentence, so DevParser's paid assessments don't guess authorship from how text reads. They run each answer against the hiring team's own hidden tests and show how it was written: editor telemetry, commit cadence, and similarity to other candidates and to public code.