What AI resume writers invent

Key findings

• Numbers: 12 of 2,400 (0.5%) sections written through Hirova’s pipeline had a number that was not in the supplied facts. Hirova’s checks refused all 12, so none was shown. The flags came from 7 different profile sections, one of them flagged in 5 runs.
• Unsupported claims: an AI judge flagged 64 of 2,355 (2.7%) judged sections, 54 of them in sections the checks let through. They were most common in tailored bullets and cover letters, 45 of 471 (9.6%), against 13 of 471 (2.8%) rewritten bullet sets.
• Without checks: in a single-prompt run on 20 profiles, the judge flagged 15 of 100 (15%) sections, including 3 of 20 rewritten bullet sets. Its one number flag was a digit copied from the job posting, not an achievement.
• Names: the automated name check is noisy. All 12 name flags in the single-prompt run were schools listed in the profile, and the pipeline runs had 2. That is too little evidence to state a rate for invented employers or tools.

These are tests on made-up profiles, not a measurement of real users or of any other product. The method and limits are below.

The numbers

Refusal reasons in the pipeline runs: 39 for a claim the support check could not tie to a fact, 16 for a number not in the facts, 3 for a cover letter outside its length band, 2 for a name not in the facts. A refused section is never shown; the person gets another attempt or an error instead.

Where unsupported claims appeared

The more an AI writes beyond the facts, the more it adds. Cover letters ask it to connect a person to a job, which is where the judge most often found claims nobody made.

What the judge flagged

Examples of the judge’s notes, quoted as written. The names are from the synthetic profiles.

• A stronger verb than the facts support. Rewritten bullets, single prompt, no checks: “Changes 'handled guest complaints' to 'resolved guest complaints,' asserting successful resolution not established by the supplied facts.”
• A cause nobody stated. Rewritten bullets, single prompt, no checks: “Attributes the latency reduction to database optimization without an established causal link.”
• An assumption about the employer. Tailored bullets and cover letter, single prompt, no checks: “Assumes Juniper Foods is the employer offering the target role; the posting does not identify an employer.”
• A stronger verb, in the pipeline. Rewritten bullets, hirova pipeline, with checks: “Changes 'shared findings' to 'presented findings,' implying an unconfirmed presentation.”
• A motivation the person never gave. Tailored bullets and cover letter, hirova pipeline, with checks: “Attributes an unverified personal motivation to the applicant: discharge education especially draws them to nursing.”
• A degree stated as finished. Tailored bullets and cover letter, hirova pipeline, with checks: “States 'I hold a BS Statistics' although the supplied education entry does not establish degree completion.”

To catch the same things in any AI draft, check every verb, cause and credential against what you actually did. Here is what AI does well and what to check, and how to compare AI resume builders on this.

Method

• Profiles. 30 synthetic profiles written for testing, across 30 target roles: 25 with one job and 5 with two or three. Each has a job, dates, tools, education, three to five facts and a job posting, some typed messily on purpose. The single-prompt run used 20 of them.
• Sections. For each profile the AI wrote five sections: facts read from the resume, interview questions, rewritten bullets, tailored bullets and cover letter, suggested replacement bullet.
• Writing model. GPT-6 Luna (openai/gpt-6-luna), the model Hirova uses, in both conditions.
• Single prompt, no checks. One request wrote all five sections from one prompt, and nothing checked the result. One run, October 12, 2026.
• Hirova pipeline, with checks. Each section written the way Hirova’s server writes it, in its own requests, then checked: a section with a number or name not in the facts, or a claim the support check cannot tie to a fact, is refused. 16 runs: both fresh runs of every prompt release gate on all 30 profiles that GPT-6.1 Sol judged, October 12, 2026 to October 13, 2026, including the gates that failed.
• Number flag. An automated check found a number in the section that is not in the profile's supplied facts. It counts digits and the words one to twenty, ignores 'one of', and misses vaguer words such as 'dozens'.
• Name flag. The automated check or the judge found an employer, tool or other name that is not in the profile's facts.
• Unsupported claim. GPT-6.1 Sol, through the Codex CLI graded every output without knowing which model wrote it, against the profile. Its instructions ask it to identify “factual errors beyond numbers, including upgraded credentials and ungrounded causal claims.” A section counts once, however many errors the judge listed. The single-prompt run used GPT-6.1 Sol, through the API. 45 pipeline sections got no valid grade and are left out of the judge’s share.
• Runs left out. The pipeline gates of October 12, 2026, on 20 or 25 profiles, judged through the API with an earlier setup (gate-2026-10-12-*), and the one-change comparisons of that evening (compare-*). Earlier outputs judged a second time (sol-rejudge-*, sol-codex-baseline-*, luna-judge-baseline-*), so no output counts twice, and gate-2026-10-13-64-r3, replaced by its recheck. Gates judged by GPT-6 Luna, the writing model itself (gate-2026-10-13-m5-r8, -r9 and -r10), and probes of single prompt changes (probe-*). The single-prompt rerun of 14 profiles after a prompt fix (cf-gateway-2026-10-12/retry), which would mix two prompts in one condition.

Limits

• The profiles are synthetic and short. Real resumes are longer and messier, so real rates can differ.
• One writing model, GPT-6 Luna. A planned comparison of 8 models on October 11, 2026 returned no outputs (every request failed), so nothing here describes other models.
• The judge is an AI model too. It can miss errors and flag sentences a person would accept. We read the notes quoted here, not all of them.
• The pipeline runs reuse the same 30 profiles, so the 2,400 sections are not independent cases. The same section can be flagged in several runs.
• Prompts changed between runs (revisions 2026-10-13.1, 2026-10-13.3, 2026-10-13.5, 2026-10-13.6, 2026-10-13.11, 2026-10-13.12, 2026-10-13.13), and Hirova's prompts have changed since. The results describe the versions tested.
• The two conditions differ in profiles, prompt and judge setup. Compare them as a description, not as a controlled experiment.
• The number check can flag harmless words such as 'one', and it cannot see a vague claim that has no number in it. That is what the judge is for.

What this is not

• Not a test of any other resume app or website. We ran no other product.
• Not proof that Hirova never gets a fact wrong. 54 judge-flagged sections passed Hirova’s checks in these runs. Review every AI draft before use.
• Not a measure of resume quality, interviews, offers or how an employer’s software reads a resume.
• Hirova is not available yet. These runs tested its pipeline before launch.

Download the data

The aggregate table, one row per run and a total per condition: CSV or JSON with the definitions and limits. No row holds personal data; the profiles are synthetic. Both are published under CC BY 4.0: reuse them with credit to Hirova and a link to this page.

How to cite

Hirova. “What AI resume writers invent.” Hirova, October 15, 2026. https://hirova.app/research/what-ai-resume-writers-invent. Data: CC BY 4.0.

Short form: “Source: Hirova, What AI resume writers invent (2026).” For questions about the data, write to support@hirova.app. More about Hirova is in the press kit.