AI resume builders make things up. Here's the fix.

Here is a small horror story, and if you have used an AI résumé tool you may have lived a version of it. A job seeker pastes in their CV. One bullet reads: “Supported the migration of internal reporting to a new analytics platform.” The tool, asked to make it stronger, returns: “Led migration of enterprise analytics platform, cutting reporting costs by 30%.”

It reads beautifully. It is also fiction. The 30% appears nowhere in the source. Neither does the “led.” The tool didn’t find those in the résumé — it composed them, because they made the sentence sound like a stronger candidate. This essay is about why that happens, what it costs you, why asking the AI nicely cannot fix it, and what actually can.

Why the machine invents

You don’t need a computer science degree for this part; the honest explanation is short. A language model is a plausibility engine. It has read an enormous amount of text and learned, with unsettling precision, what good text looks like. Ask it for a stronger résumé bullet and it knows exactly what strong résumé bullets look like: an action verb, a scope, and a number. So it produces one.

Notice what’s missing from that description: any connection to whether the number is true. The model has no access to your career. There is no database of your actual accomplishments it consults, no ground truth to check against. “Cut costs by 30%” is not a lie in the model’s world, because the model has no world — it’s simply a well-formed résumé sentence, and well-formed is what the machine is built to produce. True things and invented things can be equally plausible, and plausible is the only currency the model deals in.

This is worth sitting with, because it means fabrication isn’t a bug that better engineering will quietly fix. It’s the system working as designed, applied to a task where the design is wrong. For most writing, plausible is fine — an email, a caption, a first draft of anything. A résumé is different. A résumé is a set of claims you will be asked to defend, under questioning, by someone whose job is to probe them.

What a fabricated bullet costs you

Play the story forward. Suppose the bullet does what it was optimized to do and an interview happens. The interviewer, entirely reasonably, asks: “Tell me about that 30% — how was it measured? What was the baseline?” And now you are defending a claim you never made, in the exact moment you most need to be credible.

Your options are all bad. Improvise the details, and you have now fabricated personally, out loud, in a room. Admit an AI wrote the line, and the interview is over in every way that matters. Fumble through, and the fumbling is visible — interviewers are better at noticing a hollow claim than at noticing anything else, because probing claims is the job.

And those are just the loud failures. The quiet ones are worse because you never see them: a recruiter who knows the market reads a number that doesn’t smell right and bins you without a word. A reference check that doesn’t line up with the page. You never learn why it went silent. The cruelest part of the whole arrangement is that the cost lands entirely on you — the tool that invented the number faces no interviewer.

A résumé is a promise you have to keep in a room. Anything on it that you can’t back is not an asset. It’s a liability wearing an achievement’s clothes.

Why asking nicely cannot fix it

The obvious fix is to tell the AI to stop. “Do not invent numbers. Use only the information provided.” And to be fair, this helps — instructions like that measurably reduce how often fabrication happens. But reduce is not eliminate, and the gap between the two is exactly where the interview disaster lives. A probabilistic system gives you probabilistic obedience. A prompt is influence, not law.

Worse, the failure is silent. The model does not announce that it invented something; the fabricated version looks precisely like the honest version, one number different from the truth. So the only defense left is proofreading every generated bullet against your source, every single time — at midnight, on application thirty, which is exactly when you won’t. A safeguard that depends on tired human vigilance is not a safeguard.

There is also a logical problem underneath the practical one: you cannot prompt-test your way to a guarantee. However many times the model behaves in testing, that is evidence about the past, not a proof about the next generation. If you need a guarantee — and on a résumé, you do — it has to come from outside the model.

The fix is code, not a better prompt

The fix that actually works is almost embarrassingly old-fashioned: deterministic verification, written in ordinary code, sitting outside the model. Don’t trust the writer. Check the writing. In ScoredWell it works like this: A code check rejects any number or claim that isn’t already in your résumé.

Concretely — and I can describe this exactly, because I built it — the CV rewrite is never freeform text. The model is required to return structure: every rewritten bullet must point at the specific source bullet in your résumé it came from, by ID. No floating prose. Every line has a parent, and the parent is yours.

Then plain code — not another AI — checks each pair. It extracts every number in the rewritten bullet, in all the shapes numbers take on résumés ($3.1M, 4,200, 140%), and requires each one to already exist in that source bullet. It catches spelled-out quantities too: if the rewrite says “million” or “tripled,” the source must say it first. And pointing at a source bullet that doesn’t exist is, by itself, a violation — the model doesn’t get to invent provenance any more than it gets to invent percentages.

A violation is not “flagged for review.” The rewrite is rejected and regenerated, and if it keeps failing, the whole kit errors out rather than ships. Failing loudly is a feature. Polished fiction is the failure.

The word that matters in all of this is deterministic. The same input produces the same verdict, every time. The check can’t be sweet-talked, doesn’t hallucinate, doesn’t get tired at midnight, and costs nothing to run. Where the model is least reliable — factual restraint — it is replaced by something that cannot be unreliable. The invented number doesn’t reach the page, not because the model behaved, but because misbehavior physically can’t ship.

What the check does not catch

Honesty section, because a fix you oversell becomes its own kind of fabrication. The check verifies rewrites against your source résumé. It cannot verify your source. If your résumé says you managed a team of twelve and the truth was four, the check will faithfully preserve the fiction — garbage in, guaranteed garbage out. That part stays on you, and it should.

And the hard, mathematical guarantee covers numbers and quantity words — which is deliberate, because those are the highest-stakes fabrications, the ones an interviewer will probe first. Softer embellishment — an adjective that inflates, scope that creeps without a number attached — goes through a separate review layer that is good but is judgment, not proof. I would rather tell you exactly where the guarantee ends than imply it covers everything.

Why I care about this one

I’ve told the longer story elsewhere — a wave of layoffs, people I care about, a tool built on the evenings I still had. The part that belongs in this essay is what I watched the writing tools do: happily produce a polished CV for any posting my friends pasted in, fit or not, true or not. The polish was never the problem. The foundation was.

The promise a résumé tool should make is not “you’ll sound impressive.” It is “you can defend every line in the room.” If you take one thing from this essay, take the test: when a tool offers to write your résumé, ask what happens when the model makes something up. If the answer involves the model’s good behavior, that isn’t an answer. The fix is a checker the model can’t argue with.