AI CV Rewriting Without Inventing Facts: How to Test the Claim

Every vendor in this category says the same sentence: our AI will not fabricate. It costs nothing to say and almost nothing to believe, because the failure mode is quiet. A rewritten CV that reads beautifully and contains one invented percentage looks exactly like a rewritten CV that does not. The candidate rarely spots it. The consultant who sent it certainly does not. The client finds out at interview. This article sets out what the promise has to mean if it is worth anything, the specific ways rewrites go wrong, and the questions that separate a real guarantee from a marketing line.

Why 'we prompt the model not to fabricate' is not a safeguard

Instructing a language model not to invent facts reduces how often it does. It does not prevent it. The instruction and the output are produced by the same system, so nothing independent has checked the result. Asking a model to grade its own work has the obvious flaw: a model confident enough to invent a figure is confident enough to approve it.

The practical consequence is that prompt-only safety degrades exactly when you need it most. A well-structured CV from a senior candidate rewrites cleanly. A badly typed one, full of run-on sentences and missing dates, is where the model starts filling gaps — and that is the CV a recruiter most wants help with.

A real safeguard is deterministic, runs after generation, and can reject the model's output. It should be able to say no.

The four ways a rewrite actually goes wrong

Fabrication is not one failure. It is four, and they need different checks. Knowing them makes vendor claims much easier to interrogate.

  • Invention: a figure appears that the CV never contained. 'Improved efficiency' becomes 'improved efficiency by 30%'. The number is plausible, specific, and entirely made up.
  • Misattribution: a real figure moves to the wrong place. A £480k budget from the 2019 role appears under the 2023 role. Every number is genuine; the claim is false.
  • Inflation by omission: qualifying words disappear. 'Assisted with a £2m project' becomes 'delivered a £2m project'. No figure changed, but the candidate's role did.
  • Entity substitution: an employer, client or qualification the CV never mentioned appears in the prose, usually because the model recognised an industry and reached for a familiar name.

What a deterministic guard actually checks

ConnectIQ validates every rewrite before it reaches you, and the checks are mechanical rather than judgemental. Figures are compared per scope — the profile, each individual job, achievements, each qualification — rather than across the document as a whole. That distinction is the whole point: a whole-document check would happily let a budget move from one employer to another, because the number still exists somewhere.

Presentation is normalised before comparison, so rewriting £8m as £8 million, or fifteen as 15, is allowed. Changing £8m to £8 is not. The difference between those two cases is the difference between a formatting tool that is usable and one that rejects good work constantly.

Names are grounded too. A capitalised word in a job's bullets must come from that job or from the CV's own identity fields, so one employer cannot be name-dropped into another role's description.

There is one deliberate exception, and it is instructive. Expanding HMRC to His Majesty's Revenue and Customs adds no fact — the reader learns nothing the CV did not already say. So an expansion is allowed when the initials of the phrase spell an acronym the CV contains. That check alone is not enough: Brackenhill Construction Services spells BCS perfectly and is still an invented company name. So an expansion that passes the arithmetic is then judged on whether it is an established public body, and refused when uncertain.

What happens when the guard says no

This is the question worth asking any vendor, because it exposes whether the safeguard is real. A guard that never blocks anything is decoration. A guard that blocks work and leaves the recruiter stranded is a support ticket.

In ConnectIQ a rejected rewrite is retried free, and each retry is re-prompted with the specific rule that was broken rather than a generic warning. If a band still cannot satisfy the guard, it steps down — a full rewrite becomes a lighter one, and ultimately formatting-only, which makes no model call and always succeeds. You receive a finished CV and a note explaining what was applied.

Two consequences follow that are worth stating plainly. You are never charged for a rewrite you did not receive: attempts that fail the guard cost us tokens, not your allowance. And you never see an error where the honest answer is 'the safe version of this is a lighter edit'.

The boundary we will not overstate

Deterministic checks cover facts: figures, names, dates, qualifications, and where each belongs. They do not cover meaning. If a rewrite turns 'supported the migration' into 'led the migration', no figure changed and no new entity appeared — the checks pass, and the claim is now wrong.

That is why every rewrite is shown as a visible diff, and why the reviewer's approval is part of the product rather than an optional extra. Any vendor claiming their AI removes the need to read the output is describing something that does not exist.

The honest formulation is narrow and worth holding vendors to: invented facts are blocked mechanically; changed emphasis is caught by a human reading a diff. Both are needed.

Six questions to ask any CV formatting vendor

These are answerable in a sentence each by a vendor who has built the safeguard, and produce vague answers from one who has not.

  • Is the check deterministic, or is it the same model reviewing itself?
  • Does it compare figures per role, or across the whole document? Ask them to explain what stops a budget moving between employers.
  • What happens when the check fails — do I get an error, an unchanged CV, or a safer version?
  • Am I charged for a rewrite that failed your own check?
  • Can I see exactly what changed, field by field, before I send it?
  • What does your safeguard explicitly not cover?

Why this matters more than formatting speed

Formatting speed is a cost saving. Accuracy is a liability question. A branded CV goes out under your agency's name, and the client reasonably reads every claim on it as something you stand behind.

The tools in this category are converging on similar output quality and similar turnaround. What separates them is what happens on the CVs that do not rewrite cleanly — and that is precisely what no demo will show you, because demos use good CVs.

Frequently asked questions

Can AI rewrite a CV without inventing anything?

Not by prompting alone. It needs a separate, deterministic check that runs after generation and can reject the output — comparing figures per role rather than across the document, and requiring names to be grounded in the source. Prompting reduces fabrication; only an independent check prevents it reaching the client.

What is the difference between fabrication and exaggeration?

Fabrication introduces something the CV never contained: a figure, an employer, a qualification. Exaggeration reuses real content but changes its force — 'supported' becoming 'led'. Deterministic checks catch the first reliably. The second is caught by reading the diff, which is why the review step exists.

Does a stricter guard mean worse rewrites?

Only if it is crudely built. A guard that treats £8 million as different from £8m rejects perfectly good work and trains people to ignore it. Presentation is normalised before comparison, so equivalent ways of writing the same figure pass while genuine changes of scale do not.

What happens if the AI cannot produce a safe rewrite?

It retries free with the specific rule that was broken, then steps down to a lighter mode, and ultimately to formatting-only, which makes no AI call. You always receive a usable CV with a note explaining what was applied, and your rewrite allowance is not spent on attempts you did not receive.

Should I still read the CV before sending it?

Yes, and any vendor suggesting otherwise is overselling. Mechanical checks cover facts. Meaning is yours to approve, which is why the diff shows exactly what changed rather than asking you to re-read the whole document.

Try it on a real CV

Turn a candidate's own CV into a branded client submission you can check before it leaves the building. No card required, and every AI change is shown as a diff you approve before sending.

ConnectIQ — branded CV formatting for recruitment teams. One free conversion, no card.