What AI Gets Wrong When Rewriting a CV: Six Worked Examples
Discussions about AI accuracy stay abstract because the failures are rarely shown. That suits vendors: a fabricated CV reads perfectly, so an argument conducted in principles never has to confront one. Below are six concrete failures, each written the way it actually appears — a plausible sentence that a consultant would send without hesitation. For each, what a machine can catch, and what only a person reading the diff will.
1. The invented percentage
Source bullet: 'Reviewed the picking process and made improvements.'
Rewritten: 'Reviewed and optimised the picking process, improving throughput by 18%.'
This is the classic case, and the most dangerous because it is the most useful-looking. The model has been asked to make a bullet compelling; a specific number is the most compelling thing available, so it produces one. Nothing in the source supports 18%.
Catchable mechanically. Compare the figures in the rewritten bullet against the figures in that same source bullet: 18% appears in one and not the other, so it is rejected. Note this only works if the comparison is scoped to the role rather than the whole document — a CV mentioning 18% anywhere would otherwise let it through.
2. The budget that moved employer
Source: the 2019 role at a construction firm mentions 'fit-out packages worth £8m'. The 2023 role at a different firm mentions no figures at all.
Rewritten: the 2023 role now reads 'Delivered fit-out packages worth £8m.'
Every number here is real. The candidate did manage £8m of work. They did not manage it at the employer the CV now attributes it to, and a client checking references will discover that.
This is the failure a naive check misses entirely. If you compare figures across the whole document, £8m exists before and after, so the check passes. Conserving figures per role — profile, each job, achievements, each qualification separately — is what makes this detectable.
3. The upgraded scope word
Source: 'Assisted with the migration to a new finance system.'
Rewritten: 'Led the migration to a new finance system.'
No figure changed. No employer appeared. No date moved. Every mechanical check passes, and the claim is now materially false — this is the difference between a team member and the person accountable for a project, and it is exactly what an interview will expose.
This one is not machine-catchable, and any vendor claiming otherwise is describing intent rather than capability. Determining whether 'assisted' may become 'led' requires knowing what the candidate actually did. It is caught by reading the diff, which is why the review step exists and why removing it is not a feature.
4. The phantom employer
Source: 'Managed supplier relationships across the region.'
Rewritten: 'Managed supplier relationships across the region, including Travis Perkins and Jewson.'
The model has recognised a sector and supplied names that fit. They are real companies; the candidate may well have dealt with them. The CV never said so, and now your agency has.
Catchable mechanically, through entity grounding: a capitalised word appearing mid-sentence must exist in that role's own text or in the CV's identity fields. Names that appear from nowhere are rejected. The subtlety is that summary sections like the profile legitimately draw on the whole CV, so the rule has to be scoped rather than uniform.
5. The qualification that gained a grade
Source: 'BSc Mechanical Engineering, University of Leeds'.
Rewritten: 'BSc (Hons) Mechanical Engineering, 2:1, University of Leeds'.
Plausible, tidy, and an invented academic record. Qualifications attract this because CV conventions include grades, so a model completing a pattern supplies one.
Catchable, and best handled by not letting the model near it. Structured fields — employers, titles, dates, qualifications and institutions — are restored verbatim after any rewrite, so whatever the model returns for them is discarded in favour of the source. A rewrite should be able to change how a CV reads, never what it records.
6. The silent omission
Source: 'Reduced picking errors by about 15 percent across the night shift.'
Rewritten: 'Reduced picking errors across the night shift.'
Nothing was invented. A real, specific, favourable achievement was quietly dropped, usually because the model was tightening prose and the qualifier read as clutter. The candidate is worse represented than before you helped.
This is the case that argues against crude checking. An early version of our own guard treated any figure difference as fabrication, including omissions — which meant rewrites of badly typed CVs were rejected constantly, because parsers routinely file stray numbers like notice periods into the profile where a good rewrite removes them. Rejecting omission outright makes a tool unusable; ignoring it loses candidate evidence. The workable position is that only additions are blocked mechanically, and the diff shows removals clearly enough for a reviewer to restore anything that mattered.
What the pattern tells you
Four of these six are mechanically detectable and two are not, and the split is not arbitrary. Machines can verify whether a fact exists in the source and whether it sits in the right place. They cannot verify whether a description of someone's role is fair, because that requires knowing something the document does not contain.
That is the honest architecture of the problem, and it is worth holding vendors to it. A tool that catches none of these is unsafe. A tool claiming to catch all six is describing something that does not exist, and the claim itself should make you cautious.
The practical arrangement: mechanical checks for facts, a visible diff for meaning, and a reviewer who is expected to use it. Formatting speed is worth paying for. Unreviewed submissions are not worth accepting for free.
Frequently asked questions
How common are AI hallucinations in CV rewriting?
They cluster on poorly structured CVs, which is precisely where a recruiter most wants help. A clean, well-formatted senior CV rewrites accurately most of the time; a badly typed one with missing dates and run-on sentences is where a model starts filling gaps, because there is more ambiguity to resolve.
Can a checker catch every AI mistake on a CV?
No. Invented figures, misattributed numbers, phantom employers and altered qualifications are mechanically detectable. Changes of emphasis — 'assisted' becoming 'led' — break no factual rule while changing the claim, so they are caught by a person reading the diff rather than by software.
Why does comparing figures across the whole CV not work?
Because it misses misattribution. If a £8m budget from a 2019 role reappears under a 2023 role, the figure still exists in the document, so a whole-document check passes while the CV now makes a false claim about which employer it belonged to. Figures have to be conserved per role.
Is it a problem if AI removes a figure from a CV?
It can be, but blocking every omission makes a tool unusable — parsers routinely file stray numbers into the wrong section, and a good rewrite removes them. The workable position is to block additions mechanically and show removals clearly in the diff, so a reviewer can restore anything that mattered.
What is the safest way to use AI on candidate CVs?
Use it for structure and prose, keep structured fields such as employers, titles, dates and qualifications restored verbatim from the source, require a mechanical check on figures and names, and read the diff before sending. The speed gain survives all of those constraints; accuracy does not survive their absence.
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. 10 free conversions, no card.