Can AI write your CV without making things up?

Ask a general-purpose chatbot to "improve" your CV against a job posting and read the output carefully — genuinely carefully, line by line. There's a real chance it now says something you never did. A technology you've never touched, promoted from the posting's requirements into your experience. An accomplishment quietly inflated to rounder numbers. A responsibility that belonged to your team, reassigned to you personally.

This isn't a horror story about one bad tool. It's the default behaviour of the underlying technology, and understanding why is the difference between using AI on your CV safely and discovering the problem in an interview.

Why models invent experience

Language models are trained to produce plausible text, and their notion of plausible comes from the shape of the documents they trained on. A strong CV for a data-engineering role usually mentions certain tools. Given your CV, the posting, and the instruction "make this fit", the statistically likely completion includes those tools — whether or not they were in the input. The model isn't lying, exactly; it has no model of which claims about you are true, only of which claims are typical. When instructed to close the gap between your CV and a posting, fabrication is the shortest path, and nothing in the objective penalises it.

The failure is insidious because of where it lands. The invented material is, by construction, exactly what the posting asked for — so it reads as perfectly natural in context. You wrote the true 95 percent of the document, you're skimming your own familiar prose, and the seams don't show. The claims most likely to be fabricated are precisely the ones a reviewer is scanning for, and precisely the ones an interviewer will probe.

What it costs when it surfaces

An invented skill doesn't fail at the CV screen — it passes the CV screen. That's the problem. It fails twenty minutes into a technical interview, when someone asks a concrete question about the experience you supposedly have, in front of exactly the audience you most wanted to impress. The interviewer can't tell an AI fabrication from a candidate's lie, and the charitable interpretation isn't available to them: the CV is your document, sent under your name.

In regulated fields the stakes go up from embarrassing to disqualifying. And the reputational ledger is asymmetric — one caught fabrication colours every true claim on the page.

Grounding: the fix that's architectural, not behavioural

The unreliable fix is prompting the model to be honest — "don't invent anything" reduces fabrication without eliminating it, because the instruction is fighting the training objective. The reliable fix changes what the model is allowed to draw from.

Quarrion's tailoring is grounded in a structured profile built from your actual history — a knowledge graph of your roles, projects, skills and accomplishments, extracted from your real CV and the information you've provided. Tailoring for a posting then works by selection and emphasis: choosing which of your verified experiences to foreground, reordering, reframing vocabulary toward the posting's, trimming what's irrelevant. The generator can't claim Kubernetes experience unless Kubernetes exists in your graph, because the graph is the only place claims are allowed to come from. Fabrication isn't discouraged; it's unrepresentable.

That's the distinction worth taking to any tool in this category: does it generate from your verified history, or generate plausible text and hope? The first architecture makes honesty structural. The second makes it a behaviour you have to audit for, forever.

What honest tailoring still does

Ruling out invention doesn't reduce tailoring to formatting. Legitimate, grounded tailoring still moves the needle: surfacing the two genuinely relevant projects that were buried under chronology; translating your vocabulary to the posting's where they name the same skill; cutting the half-page that means nothing for this role; leading with the achievement this hiring manager will care about. All of it is your real experience, arranged for this reader — which is what tailoring was supposed to mean before the shortcut became available.

The gap between your experience and a posting's wishlist, where it's real, is information. Sometimes it says "apply anyway, and be ready to talk about the adjacent thing you did instead" — a case honest materials let you make credibly. Sometimes it says this role isn't the one, and the evening belongs to a better-matched application. A tool that papers over the gap destroys the signal both ways.

The test worth running

Before trusting any AI CV tool with something sent under your name: feed it a posting requiring a skill you demonstrably lack, and see what comes back. A grounded system works with what you have or leaves the gap visible. A plausibility engine gives you the skill. It's a two-minute experiment, and it tells you which kind of ghostwriter you've hired — one constrained by your record, or one constrained by nothing.