Honest AI

Can AI Invent Experience on Your Resume?

Most AI resume tools will happily write a metric you never earned. Here is why they do it, what it costs you at interview, and what a fabrication guard checks.

Yes. Most of them will, and they will do it without any warning that something has been made up.

Ask a general-purpose AI writing tool to "make this resume stronger for this job" and you will often get back a document that reads better than the one you gave it and describes a slightly different person. A 22% cost reduction appears where you wrote "reduced costs". Kubernetes appears in a skills list because the posting asked for it. A job title creeps upward by one rung. None of that is malice. It is the predictable output of the machinery.

Why a language model writes a number you never earned

A language model is trained to continue text plausibly. It has a very good internal model of what a persuasive resume bullet sounds like and no model at all of what happened in your career. When you instruct it to make a line stronger, it reads that as make this sound stronger, and the cheapest way to make a claim sound stronger is to attach a figure to it.

So "improved reporting speed" becomes "improved reporting speed by 40%". The 40 is not a guess about your work. It is a guess about what number usually sits in that slot in the training data. The same mechanism fills in tools, team sizes, budgets, client names and years.

This gets worse, not better, as the tool tries harder to help. A model told to close the gap between your resume and a demanding job description has exactly two ways to do it: surface things you already have, or produce things you do not. The first is bounded by your history. The second is not. Under pressure to raise a score, the unbounded option always wins.

The four shapes invention takes

Fabrication in a tailored resume is not one failure. It is four, and they are worth naming because you catch them differently.

The last one is the quiet killer. Everything in the sentence is real. Only its placement is false. A human proofreading their own tailored resume very rarely spots it, because every component is familiar.

What it costs you in the room

An interviewer who reads "reduced processing time by 40%" will ask what it was before, what it was after, how you measured it and who else was involved. That is a normal, friendly question. It is also unanswerable if the number arrived from a model.

You will rarely be caught in an outright lie at this stage. What happens instead is that you are caught being vague about your own headline achievement. You hedge, the interviewer's attention sharpens, and the rest of the conversation is spent testing you rather than getting to know you. A real but unremarkable claim you can narrate for four minutes is worth more than an impressive one you cannot survive two questions on.

What it costs you after the offer

Verification is not a rare event. HireRight's 2025 Global Benchmark Report found that more than three-quarters of employers surveyed uncovered a discrepancy on a candidate during the previous twelve months, and that employment verification was the most common source of them across every region measured.

Employment verification checks titles, dates and sometimes salary against the employer's own records. It does not care how the discrepancy got there. A title inflated by a tool you trusted looks identical, in a screening report, to a title you inflated yourself. The consequence typically lands after you have resigned from your current job, which is the worst possible moment for it.

What a fabrication guard actually does

A fabrication guard is a check that runs after generation and before anything reaches you. It treats the model's output as a suspect rather than a product.

The mechanism is token traceability. Every meaningful token in the tailored document has to trace back to a source in the original resume. Where it cannot, the guard rejects it. In Involve, four categories fail the check:

1. A number with no source. Any figure, percentage, currency amount or duration that does not appear in the original document.
2. An entity with no source. An employer, organisation, client or product name absent from the original.
3. Any other token with no source. A tool, language, certification or qualification the original never mentions, however well it would fit the posting.
4. A phrase whose evidence sits on a different line. Content that exists in your resume but has been recombined, so an achievement from one role now sits under another.

The important design decision is what happens next. The guard throws rather than ships. A tailored document that fails the check does not get quietly trimmed and handed to you with the offending line removed; the tailoring fails, and you are told. Silent repair is how fabrication survives, because a shorter document raises no questions.

The guard runs over the resume, in your own browser, alongside the parsing, the scoring and both file writers. Nothing about the check depends on you noticing anything.

What it does not do

This is where honest tools have to be honest about themselves.

A fabrication guard verifies the tailored document against your original resume. It does not verify your original resume against reality. If you wrote "led a team of eight" and it was three, the guard will carry that through without comment, because the original is its only reference point. The guard constrains the machine, not you.

It also does not make a weak application strong. Constraining a tool to your real history means some jobs will come back with a poor score, and that score is information. Involve reports it as two separate numbers: Fit, which asks whether your background matches the role, and CV readiness, which asks whether the document survives the screen. A low Fit will not improve because a model added a keyword. It improves when you find a role your history actually supports, or when you do the work the role requires.

What honest tailoring leaves you with

Plenty, as it turns out. Reordering bullets, reweighting them, retargeting the summary and rewriting weak openers verb-first changes what a reader sees in the first eight seconds without changing a single fact. Most resumes bury their best evidence in the third bullet of the second role. Moving it is not deception; it is editing.

Every change Involve makes arrives as a before-and-after diff with a reason attached, and needs your explicit approval. Your base resume is never overwritten, and each tailor is a new document, so you can always see what the original said. If you want the full picture of how the pieces fit together, start here, or open the app and run your own resume through it.

The test for any AI resume tool is simple. Give it a thin resume and a demanding posting, and read what comes back for things you never wrote. If they are there, the tool is optimising for a good-sounding sentence, and you are the one who has to sit in the room and defend it.