Can Your Resume Formatter Invent a Job Your Candidate Never Had?
There is a lot of writing about candidates using AI to embellish resumes. There is almost none about the tool sitting between the candidate and your client doing the same thing on its own initiative. That is the more dangerous direction, because a candidate's exaggeration is the candidate's, while a fabrication introduced during formatting is yours — it arrived on your letterhead, under your right to represent, after you pressed the button. This is what that failure actually looks like, and how to test for it in about ten minutes on any product including ours.
Why fabrication is hard to see on a resume
A language model asked to improve a resume is optimising for prose that reads well. Nothing in that objective distinguishes a fact it was given from a fact it supplied. The output is fluent either way, and fluency is exactly what stops you noticing.
It is also the wrong shape for human review. A recruiter reading a submittal is checking whether the candidate looks right for the role. They are not diffing against a document they read twenty minutes ago. Two extra words in a bullet — "managing a team of twelve" where the original said "managing a team" — survive that reading every time.
And the error is invisible from the client's side too. Your client cannot tell an invented figure from a real one. They find out at interview, when the candidate cannot substantiate something they never claimed.
The four failure modes, in the order they cost you
These are not hypothetical categories. They are what we found when we deliberately tried to break our own rewrite pipeline, and each one needs a different control.
- Invented employer or institution. The model expands "BCS" into a company name that spells BCS perfectly and does not exist. Initials matching is not evidence.
- Drifted date. A role written as "2019 to 2022" comes back "2019 to 2023", usually because the model normalised an adjacent range and carried the wrong end.
- Changed or moved figure. The most expensive one. A number that existed in the source moves to a different job, attaches to a different metric, or reverses direction: "revenue grew 20% and costs fell 10%" becomes "revenue grew 10% and costs fell 20%". Every figure is real. The claim is false.
- Phantom skill. The model adds a technology that fits the role profile and appears nowhere in the candidate's history. This is the one candidates get caught on in a technical screen.
The ten-minute test
Run this on a real resume you have already submitted, not a sample the vendor gives you. Vendor samples are chosen.
Take a resume with at least three roles and several numbers in the bullets. Run it through the tool at its strongest rewrite setting. Then check the output against the original in this order — it takes about ten minutes and it will tell you more than any demo.
A product that passes all six is not necessarily safe, but a product that fails any of them is telling you something specific about how it was built.
- Count the numbers. Every figure in the output should exist in the input. Not "roughly the same" — the same.
- Check which job each number is attached to. This is where tools fail most often, and it is the failure a reviewer is least likely to catch.
- Check direction words. Grew, fell, reduced, increased. A preserved number with a reversed verb is a reversed claim.
- List the employers and schools in the output and find each one in the original. Any expansion of an acronym deserves particular attention.
- Scan the skills for anything the candidate never mentioned.
- Now delete a section from the source — remove the education block entirely — and run it again. A tool that fills the gap with something plausible has just told you it will do that on every incomplete resume you upload.
The only control that actually works
Everything above is detection. The control is that a rewrite has to be checked mechanically before you ever see it, and that the check has to be able to refuse.
In our own pipeline that means structured fields — employers, titles, dates, qualifications — are restored word for word from the parse rather than trusted from the model. Figure tokens have to form the identical set within each scope: within one job's bullets, within the profile, within one qualification. A number cannot move between jobs, because the comparison is per section rather than per document. Common polarity reversals are rejected. Capitalised entities that were not in the source are rejected outright.
When a check fails, the rewrite is retried with the rule it broke, free, and if it still cannot pass, the intensity steps down until it reaches formatting-only, which makes no model call and always succeeds. You get the resume back with an honest notice rather than an error, and you are never charged for an attempt that failed.
The part we cannot mechanise, and say so: whether the rewritten prose still means what the candidate meant. That is a judgement, it is shown to you as a diff, and you approve it. Any vendor claiming to have automated that has not thought about it hard enough.
Why this is a right-to-represent problem, not just a quality problem
A submittal is a representation. You are telling a client this person has this history. The RTR you took from the candidate covers representing them; it does not cover representing a version of them that a tool composed.
If a fabricated figure reaches a hiring manager and the candidate cannot support it at interview, the damage lands in the order you would least want: the candidate looks dishonest, you look careless, and the client's next three requisitions go to someone else. None of those outcomes are recoverable by explaining that the AI did it.
The asymmetry is the argument for a boring answer. A blank field is a nuisance. An invented one is a liability. Every design decision in this area should follow from that, and if a vendor's does not, the resulting product will be more impressive in a demo and worse on your desk.
Frequently asked questions
Doesn't a good prompt stop AI from making things up?
No. Prompt instructions reduce the rate; they cannot enforce it, because the model has no mechanism to check its own output against the source. Anything that matters has to be verified in code after generation, with the ability to reject the result. Ask a vendor whether their guarantee is a prompt or a check — the answer is usually in how quickly they change the subject.
What is the single most common fabrication in resume rewriting?
A figure that exists in the source attaching itself to the wrong role or the wrong metric. It defeats naive checking because every number is genuine — only the association is wrong. Catching it requires comparing figures within each section rather than across the whole document.
How do I test a vendor without uploading real candidate data?
Use a resume with your own details in place of the candidate's, keeping the structure, dates and figures intact. The failure modes are about structure and numbers rather than identity, so a substituted name and contact block changes nothing about what you will find.
Is a tool that refuses to rewrite sometimes a worse tool?
It is a more honest one. A rewrite that cannot pass a factual check has two possible outcomes: return the source with an explanation, or return the unverified output. Only one of those keeps an invented employer off a client submittal. Our pipeline retries free, then steps the intensity down rather than dead-ending.
Should we tell candidates their resume was AI-rewritten?
It is good practice and increasingly the direction of travel in state law — Illinois HB 3773 has required notice for AI use in recruitment since January 2026, though its target is decision-making rather than document formatting. Separately, showing the candidate the reformatted version before it goes out catches errors no automated check can, because they are the only person who knows what actually happened in the job.
Try it on a real resume
Turn a candidate's own resume into a branded client submittal 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 resume formatting for recruitment teams. 10 free conversions, no card.