AI Resume Tools Are Inventing Your Experience. Catch It First

By , founder of CVBooster · Published · Updated

10 min read

A hiring manager opens with an easy question: "Walk me through the Kubernetes migration you led." You never led a Kubernetes migration. You are not sure what one involves. But there it is, third bullet of your most recent role, because twenty minutes before the deadline you pasted a job description into an AI resume tool, clicked optimize, and shipped what came back. This is the unsolved problem in the category, and the invented lines are almost always the most impressive ones on the page.

Why models fabricate when you ask them to improve a resume

A language model produces text that is plausible given the text it has already seen. Plausible is not true, and the model has no mechanism for telling them apart. Hand it your resume plus a job description and you have given it two documents: one describing what you have done, one describing what somebody wants. The gap between them is the only interesting thing in the prompt. To you that gap is a fact about your life. To the model it reads as a defect in the draft.

The most plausible continuation of "resume for a candidate applying to a role requiring Terraform" is a resume that mentions Terraform. So it writes one. Nothing penalises that: there is no verification step and no external record it can consult. Everything you did not type into the box does not exist as far as the model is concerned, and everything the posting asks for is sitting in context, repeated four times, clearly important.

Verbs make it worse. Optimize, strengthen, improve and tailor all imply a direction of travel, and the only direction available is toward the job description. Ask for quantified bullets and you are asking for numbers, and if you did not supply any the model will, because a bullet with a figure reads as stronger than one without.

Tip: Test any tool before you trust it. Feed it a deliberately thin resume, three lines, no metrics, plus a demanding job description. If what comes back is full of confident figures and tools you never mentioned, that is not tailoring. It is a fictional candidate sharing your name.

The five failure modes, in order of how often they bite

1. Certifications that were never earned

The most dangerous one, because certifications are trivially easy for a model to pattern-match. The posting says AWS Solutions Architect preferred, your resume mentions cloud work, and the rewrite quietly gains a line under a heading called Certifications. Sometimes with a plausible year attached. Sometimes as "AWS certified" buried mid-sentence in the summary, where you will not catch it on a skim. Certifications are also the most verifiable claim on a resume, which matters later.

2. Tools you have heard of but never touched

Your skills line said SQL and Excel. It returns as SQL, Excel, Tableau, Looker, dbt and Snowflake, because those words cluster together in training data and in the posting. The bullets follow: "built automated reporting in Tableau" reads well and describes something you did in Google Sheets. This one is subtle, because the underlying achievement is real. Only the instrument is wrong, and the instrument is what a technical interviewer asks about.

3. Metrics conjured out of nothing

"Improved process efficiency" becomes "improved process efficiency by 35%." You never measured it. Nobody measured it. The 35% is decoration, and it invites a follow-up: measured how, over what period, against what baseline, who else saw the number. Invented metrics turn a conversation into an interrogation, because a fake figure is easy to probe and hard to defend without visibly improvising.

4. Real achievements smoothed into nothing

This one is quieter and probably costs more interviews than the rest combined. You wrote something specific and slightly odd: "rebuilt the night-shift handover checklist after two medication errors, incident reports dropped to zero the following quarter." The rewrite returns "implemented process improvements to enhance operational quality." Nothing false was added. Your most interviewable, most obviously human line was replaced with a sentence that could belong to anyone. Models regress toward the mean of their training data, and that mean is corporate mush.

5. Title and scope inflation

Coordinator becomes manager. "Supported the rollout" becomes "led the rollout." "Worked with the vendor" becomes "owned the vendor relationship, negotiating a 20% cost reduction." Team sizes appear that you never supplied. This feels like the most forgivable category because the work was genuinely yours, but titles get verified in reference checks routinely, and led versus supported is exactly the distinction a good interviewer spends ten minutes pulling apart.

The part nobody writes about: you have to defend this document out loud

Resume advice usually stops at submission, as if the goal were the interview rather than the job. That framing is what makes fabrication feel cheap. It is a loan against a conversation you have not had yet.

Behavioural interviewing treats your resume as a list of prompts. The interviewer picks a line, usually the most impressive one, asks you to expand, then asks a follow-up you cannot anticipate because it is generated from your answer rather than a script. What was the hardest part. Who disagreed and what was their argument. What broke first. What did you try before the thing that worked.

Real memories survive that indefinitely, because a real project has infinite detail and you were there for all of it. Fabricated lines collapse in about ninety seconds, and they collapse recognisably. Answers get more abstract instead of more concrete. Specifics arrive only after a pause. You describe what one would generally do rather than what you did. Anyone who has run fifty interviews knows the difference between recalling and constructing.

The outcome is worse than a plain rejection. You are not scored as underqualified. You are scored as someone whose account of themselves did not hold up, and that note follows you through the debrief and often into later applications there.

And if it survives the interview, there is still the background check

Employment dates, titles and degrees are routinely verified by third-party screening firms before a start date. Certifications are worse, because most major issuers publish a public verification registry. PMI, AWS, Microsoft, Google Cloud, CompTIA, Cisco and the main accounting bodies all let anyone confirm a credential in seconds. A fabricated certification is not a risky claim. It is one that can be disproved by a recruiter with a browser tab open.

Discovery before the offer costs you the role. Discovery after you start is a different category: most contracts treat material misrepresentation in the application as grounds for termination, frequently without notice and outside the usual performance process, and it can resurface years later during a promotion or a clearance. That is the real risk profile of a line a tool wrote in two seconds.

Tip: The line you are most tempted to keep because it is doing so much work for you is, statistically, the invented one. Impressive and unverified is not a coincidence.

The audit: four passes, about twenty minutes

None of this needs a tool. It needs you to read your own resume as an adversary rather than an author, every time you accept AI output, including output that came back looking modest.

Pass one: the two-minute story test

Take each bullet and talk out loud, to an empty room, for two minutes about it, without repeating the bullet's wording. Say who asked for the work and why, what you tried first, what went wrong, who pushed back, how you knew it worked. If you dry up inside thirty seconds, one of two things is true. Either the line is not yours, so delete it, or it is yours but written so vaguely it triggers no memory, so rewrite it using the specifics that surfaced while you struggled. This pass catches most fabrication and doubles as interview preparation.

Pass two: every number gets a source

Highlight every digit in the document. For each one, name the artefact it came from: a dashboard, an invoice, a ticket queue, a headcount, a report you emailed on a specific Friday. "I remember it being about that" is not a source. If you cannot name the artefact, cut the number or replace precision with an honest range. "Cut roughly a third of the manual steps" is defensible. "Reduced manual effort by 37%" is defensible only if you can say where 37 came from, and false precision attracts exactly the follow-up you cannot answer.

Pass three: every tool, system and standard

List every proper noun naming software, a framework, a methodology or a regulation. For each, answer two questions: what did you personally use it for, and what annoyed you about it. The second is the useful one. Genuine use always generates complaints, because everything is irritating once you have used it under deadline. If a tool produces no opinion in you, that is recognition rather than experience, and recognition does not survive a technical screen. Cut it, or demote it to a labelled "familiar with" line that claims no delivery.

Pass four: read the diff, not the output

Keep the version that existed before the AI touched it and compare the two side by side. Most people read only the new document, so they see what was added and never notice what was removed. Ask two questions of the diff: what claims exist now that did not exist before, and what specific, strange, true detail got smoothed into a generic phrase. Restore the second category by hand. The weird specific detail is what an interviewer remembers three days later.

Where the real line sits between rewriting and fabrication

None of this argues against using AI on your resume. Refusing all machine help because some of it lies is an overcorrection that costs you interviews. The distinction is easy to apply once you see it. Rewriting changes emphasis, ordering and vocabulary. Fabrication changes the set of claims. Ask of every edited sentence: does this assert anything that was not true before the edit? If not, it is rewriting, however much better it sounds.

Legitimate, every time:

Fabrication, every time, regardless of how the tool phrases it:

There is one honest grey zone: shared credit. Real work is collaborative and resumes are individual documents, so some claiming is unavoidable. The workable rule is that you can claim the outcome if you can name your specific contribution without flinching. "Led the migration" is fine if you owned the plan and the sequencing, even though eight people did the work. It is not fine if you attended the standups.

What to do about a gap that is actually real

People accept fabricated lines not out of dishonesty but because the alternative feels like giving up. The posting wants five years of Terraform and you have none, so the choice looks binary. It is not. Postings are wish lists assembled from several people's opinions, and hiring teams constantly interview candidates who miss listed requirements, because the alternative is interviewing nobody.

You have four honest options. Name the closest adjacent thing you genuinely did and let the interviewer judge the transfer. State learning in progress accurately, describing what you have completed rather than what you intend to. Address it in one sentence of the cover letter so it stops being a discovery. Or decide the gap really is disqualifying and put the hour into a posting where it is not. All four leave you able to speak freely. A papered-over gap does the opposite: it makes the topic you most need to handle well into the one you cannot discuss.

Tip: A gap you name yourself becomes a data point. A gap the interviewer finds becomes a credibility problem. The information is identical. Only the sequencing differs, and the sequencing is yours to control.

What to demand from any AI resume tool

Judge tools on this, not on how confident the output sounds:

Be wary of pure score optimisation. Much of the category sells a number that rises: Jobscan is $49.95 a month with five free scans, Teal $29 a month, Rezi $29 a month. Those scores largely reward keyword presence, which creates direct pressure to add keywords, which is the fabrication path with a progress bar attached. A score climbing from 62 to 91 is not evidence that anything true about you improved, only that the document now contains more of the posting's words.

CVBooster's tailoring is built around the refusal rather than the fill-in. Where the posting asks for something your material does not support, it is meant to surface that as a gap you decide about rather than quietly write a sentence covering it. Less satisfying in the first thirty seconds, considerably better in the interview. Either way, the audit above is yours regardless of which tool you use, or whether you use one at all.

Looking for a worked example for your job title? Browse the resume examples by job title.

Get your free ATS score · All articles

Built by Mustafa Tarabya at DT Nova. CVBooster for iPhone is on the App Store.