AI résumé tools that do not invent experience
Hallucinated experience is the one AI résumé failure you cannot talk your way out of. How the safeguards work, and how to test for them.
There is one way an AI résumé tool can hurt you rather than just waste your time: it writes something you did not do, you do not notice, and a hiring manager asks about it.
Everything else — clumsy phrasing, a bad template, a score that means nothing — costs you an afternoon. This one costs you the interview, and sometimes the reference.
This page is about that failure mode specifically: why it happens, what a safeguard actually looks like, and how to test a tool for it before you trust it.
Why tools invent experience
A language model asked to make a résumé match a posting will close the gap the cheapest way it can, and the cheapest way is words. If the posting wants Kubernetes and your résumé does not say Kubernetes, adding the word is a smaller edit than explaining why the rest of your container experience transfers.
Three product decisions make it worse:
- Scoring the output against the posting. If the tool grades itself on keyword overlap, invention is the shortest route to a better grade.
- Rewriting the whole document at once. A wholesale rewrite hides the single changed word in a wall of plausible text.
- No per-edit approval. If accepting means accepting everything, nobody reads everything.
The invention is rarely dramatic. It is "led" where you wrote "worked on". It is "reduced costs by 30%" attached to a project that had no cost measurement. It is a skill in a skills list. These survive a skim, which is exactly the problem.
What a real safeguard looks like
"Our AI is accurate" is not a safeguard. A safeguard is a check that runs on the output and removes things, and that you can describe in one sentence. Four that are worth asking about:
Source anchoring. Every suggested edit has to quote an exact passage from the résumé you uploaded. If the quoted original does not appear in your file word for word, the suggestion is discarded rather than shown. This kills the whole class of edits that invent a new bullet point, because a new bullet has no original to quote.
Numeric grounding. A replacement may not contain a figure that was not already in the passage it replaces. This is crude — it cannot tell a good metric from a bad one — but it mechanically removes the invented percentage, which is the most common and most damaging fabrication.
No overlapping rewrites. If two suggestions touch the same sentence, applying both produces a sentence neither of them proposed. Dropping the overlap keeps every accepted edit meaning what it said it meant.
An untouched original. The file you uploaded stays as you uploaded it. Tailoring produces a separate set of proposed changes, so a bad session costs you nothing and does not compound into the next one.
Notice what none of these do: they do not verify that a true statement is a good statement, and they cannot catch an invented claim that contains no numbers and happens to quote a real sentence. A safeguard narrows the failure. It does not remove your review.
Testing a tool in ten minutes
Take your real résumé. Pick a posting asking for one genuine gap — a tool, a language, a scale of team you have not worked at. Run the tailoring. Then check, line by line:
| Check | What you are looking for |
|---|---|
| Unsupported skills | Anything in the skills list that is not in your original |
| Invented numbers | Percentages, headcounts, revenue, latency figures that appeared |
| Seniority drift | "contributed to" → "led", "helped" → "owned", "part of" → "managed" |
| Date and title edits | Any change to employers, titles, degrees or date ranges |
| Promotion invention | A role split into two, or a title you never held |
| Recoverability | Can you still open your original file, unmodified? |
A tool that fails the first two is not usable for this. A tool that fails only the last is usable but you should keep your own copy.
Run the same test on the cover letter. Cover letters have fewer structural constraints and are where invention tends to hide.
What SchoolWhool does
SchoolWhool tailors an existing résumé; it does not build one from scratch. When you run Analyze & tailor on a role, the model is instructed never to invent employers, experience, education, metrics, skills or work authorisation, to put missing qualifications in a gaps list rather than in the résumé, and not to upgrade seniority. Instructions alone are not the safeguard. Three checks run on what comes back:
- Each edit must quote an exact, unique contiguous passage of your uploaded résumé. Not found, or found twice? Discarded.
- Edits whose quoted passages overlap are dropped.
- Any replacement introducing a number absent from the passage it replaces is dropped.
You then approve edits one at a time. Your original résumé is retained. Changing your profile details invalidates an earlier approval, so a review never silently carries over onto different facts. Approving materials does not submit an application — you apply on the employer's own site.
The fit assessment that comes with it is SchoolWhool's own reading of the posting against your profile. It is labelled that way throughout the product because that is what it is: not an employer ATS score, not a hiring prediction.
Honest limits. These checks stop invented numbers and invented bullets. They do not stop a rephrasing that is technically anchored to your text but overstates it, and they do not judge whether a claim is wise to make. Read the edits.
If you only remember one thing
Before you pay for any AI résumé tool, give it a gap and read what it writes. Ten minutes on a free tier tells you more than any comparison table, including this one.
Browse live roles free, or read how to tailor a résumé to a job description if you would rather do it by hand.