The Resume Hallucination Report: How Often AI Invents Things on Your Resume (2026)

The short version: AI resume tools share a quiet problem — in the process of “optimizing” your resume for a job, the AI can add things you never did. We measured it. In our internal testing, we tracked 106 AI-tailored resume bullets across 8 professions, checking every rewritten line against the candidate’s source resume, and identified the 5 specific ways AI inflates a resume. With grounding constraints and a second verification pass, we observed zero fabrications — 0 invented tools, 0 inflated titles, 0 made-up partnerships. Here’s the data, the method, and why “defensible” matters more than “optimized.”

0 fabricated claims observed in 106 AI-tailored bullets, across 8 professions — with grounding + verification. Unconstrained, AI often goes beyond stretching the truth into hallucinated titles and experience. It’s one thing to reword a job title from “Customer Service Rep” to “Customer Service Associate” — it’s another to inflate it to “Client Consultant.” Bloom’s approach is user-first: tailor the resume to the job, then verify every line against your real experience, so the framing stays defensible.

What “resume hallucination” actually means

When an AI rewrites a resume bullet to match a job description, it can introduce claims the candidate’s real experience doesn’t support: a tool they never used, a title one level higher than reality, a cross-team partnership that never happened. It isn’t lying on purpose — the model is optimizing for “sounds like a fit for this job,” and the fastest way to sound like a fit is to borrow language straight from the job posting, whether or not it’s true of you.

That’s the trap. The resume scores better and reads stronger — right up until a recruiter asks you to talk about the thing the AI invented.

The 5 ways AI inflates a resume

Across our tests, fabrications clustered into five recurring patterns. If you use any AI resume tool, these are the five things to check before you hit send:

Invented domain qualifiers. Words like “data-driven,” “self-serve,” “high-impact,” or “GTM-focused” get added when the source never used them. In our spot-checks this was the most persistent pattern — qualifiers feel harmless, but they’re claims.
Invented tools and platforms. A resume that mentions “a dashboard” comes back claiming “Tableau,” “Looker,” “dbt,” or “Snowflake” — tools pulled from the job description, not from your history.
Title and seniority promotion. “Co-managed a budget” quietly becomes “owned” or “led.” A “marketing operations leader” gets reframed as a product manager.
Invented partnerships. “Partnered with the sales engineering team” appears when no such collaboration exists in the source.
Promoted certainty. “Contributed to” becomes “led.” “Supported” becomes “drove.” The work is real; the level of ownership claimed is not.

What we found

We measured the problem in two stages. First, in earlier rounds of our evaluation series, we ran an unconstrained tailoring prompt — the default way most tools use AI — repeatedly on the same resumes and job descriptions. Then we ran grounded tailoring plus a separate verification pass that checks every bullet against the source resume, across the full 106-bullet set.

Scope: 106 tailored bullets across 8 professions — software engineer, product manager, data analyst, career switcher, registered nurse, licensed electrician, federal analyst, and hospitality manager — over 12 resume-to-job pairings, each pairing run 4 times.
Unconstrained AI was unstable, and it invented. Run the same unconstrained prompt on the same resume repeatedly and the per-bullet grounding verdict — is this line fully supported by the source resume? — flipped between runs on roughly 40–50% of bullets, depending on the fixture set. In our sample, about three-quarters of those flips were wording drift rather than new facts. The rest was the real problem: genuinely fabricated material — invented tools and unearned qualifiers — at just under one fabricated fact per run on average.
With grounding + verification: 2 verdict flips in 106 bullets (1.9%) — and both traced to the checker’s limited view (one waffled on a percentage correctly derived from the source; one flagged a summary that drew on the Skills and Education sections it can’t see), not to the generator inventing anything. Zero invented tools, titles, or partnerships were observed across all 106 bullets.

Why this matters: defensible AI, tailored to you

The real giveaway of AI on a resume isn’t bad writing — it’s quiet inflation you can’t back up. A resume that scores well but lists a tool you’ve never opened is a liability the moment the interview starts.

So the bar we hold ourselves to isn’t “a resume that beats the ATS.” It’s defensible AI, tailored to you: a resume tailored to the specific job, then verified against what you actually did, with every bullet labeled supported or not — so the version you send is one you can stand behind in the room. That’s the difference between a resume that gets you the interview and one that survives it.

We built this report without touching a single user’s resume

This study ran entirely on fictional resumes — invented candidates across 8 professions — not customer data. That’s deliberate, and it reflects how Bloom treats user information:

Bloom does not sell user data. Our published Privacy Policy says it plainly: “We do not sell your personal information, and we do not share it for cross-context behavioral advertising.”

Your resume is yours. Bloom has no employer-facing side — sharing is limited to the service providers that run the product, and nothing you write in Bloom is shown to employers.
This research used no user data at all. Every resume in the study was fictional, so there was never anything to anonymize in the first place.

Privacy-first isn’t a footnote here; it’s written into our Privacy Policy and Terms of Service.

How Bloom catches it

Two ideas do the work:

Constrained generation. The AI tailors your resume under explicit rules: no tools or platforms you didn’t list, no promoting “co-managed” into “led,” no unearned qualifiers, no invented partnerships. It can rephrase, reorder, and tighten what’s true; it can’t add what isn’t.
Independent verification. A separate pass then checks every tailored bullet against your source resume and labels it — supported, with the exact span of your experience that backs it, or flagged, with the reason.

You see the result per bullet, so you’re never trusting the AI blind. More on the philosophy in Using AI on your resume — honestly and How to use AI without lying on your resume. See how it works.

How we measured it

All figures come from our internal May 2026 grounding evaluations. We built fictional candidate resumes across 8 professions, paired them with realistic job descriptions (12 resume-to-job pairings), and ran each pairing through the tailoring pipeline 4 times. Every tailored bullet was checked against its source resume — by the verification pass and by manual review — and we counted a bullet against the system if its supported verdict was inconsistent between runs or if it contained material not traceable to the source. The unconstrained baseline comes from earlier rounds of the same evaluation series on overlapping fixtures. These numbers describe what we observed on our test set; they’re an honest measurement, not a guarantee.

FAQ

Do AI resume builders lie?

Not deliberately — but unconstrained AI tailoring optimizes for matching the job description, and it will happily add claims that score well without checking them against your real experience, skills, education, and certifications. In our tests, unconstrained tailoring was unstable from run to run and occasionally invented tools and qualifiers outright. Grounding rules plus a verification pass drove observed fabrications to zero.

Can a recruiter tell if my resume was inflated by AI?

Often not from the document itself — but the interview is where invented tools, titles, and partnerships fall apart. The risk isn’t detection software; it’s being asked to talk about something you didn’t do.

Is it safe to use AI on my resume?

Yes — if the tool grounds its output in your real experience and verifies each claim. Unverified AI tailoring is where the risk lives. See Is it OK to use AI on your resume?

How was this measured?

On fictional resumes across 8 professions and 12 resume-to-job pairings, each run repeatedly through AI tailoring, we checked whether every tailored bullet could be traced back to the candidate’s source resume. The unconstrained baseline was measured the same way in earlier rounds of the series. The “How we measured it” section above has the details.