Why do AI resume tools make up numbers?
Because the model is asked for a bullet that sounds like a strong resume, and strong resumes have numbers. When your input has none, the model fills the gap with a likely-looking figure. A 2025 paper by researchers at OpenAI and Georgia Tech argues models guess like this because training and testing reward a guess over admitting they don't know.
The paper, Why Language Models Hallucinate by Kalai, Nachum, Vempala and Zhang, compares the model to a student facing a hard exam question. Under the usual scoring, a blank or "I don't know" earns nothing and a guess sometimes earns a point, so guessing wins. Your resume is that exam, and "by how much?" is the hard question.
The US National Institute of Standards and Technology calls the same thing confabulation. Its Generative AI Profile, NIST AI 600-1, published in July 2024, defines it as confidently stated but false content. NIST says it is a natural result of how these models work: they produce text that matches the patterns in their training data.
Why do resume tools reach for numbers in the first place?
Because nearly all resume advice asks for them. Amazon's own interview and resume guide, published April 8, 2025, tells candidates to show the impact of their work. Its example turns "managed a team's cloud migration project" into a bullet with a 15-person team, a two-month lead and $1.2M in annual savings.
That advice is sound when the numbers are yours. If the resumes and advice a model learned from look like Amazon's example, the pattern it picks up is that a good bullet ends in a figure. Ask it to "make this more impactful" and it will supply one, whether or not it exists.
The figure tends to be round and plausible: 20%, 30%, 40%. Nothing in the output marks it as invented. It sits in the same font as your real dates and job titles.
What does an invented number look like?
Here is a worked example of an invented 40%. The bullets are written for illustration and are not output from any named tool. A candidate gives a general AI writing tool one plain, true line and asks it to add metrics. The model returns a number that sounds measured, though nobody measured it.
| Version | Bullet | Where the number came from | Survives "how did you measure that?" |
|---|---|---|---|
| Your original | Ran the weekly support rotation for the payments team | No number | Yes, there is nothing to defend |
| AI rewrite, no facts given | Cut ticket resolution time by 40% by redesigning the weekly support rotation | Invented by the model | No, you never measured it |
| Rewrite after you confirm a fact | Ran the weekly support rotation for the payments team, closing about 60 tickets a week | Your own ticket dashboard (for example) | Yes, you can say where it came from |
| Rewrite with no number available | Ran the weekly support rotation and wrote the handover checklist the team still uses | No number; a result someone can check | Yes |
Look at what changed in row two. "Ran" became "cut," a cause appeared ("by redesigning"), and 40% arrived to measure it. Three claims were added, and you made none of them. The 60 tickets in row three is also just an example; the point is that it comes from a source you can name.
Why is a made-up number worse than no number?
A made-up number is a claim an interviewer can test, and you can't pass the test. Numbers on a resume invite follow-up questions: what was the baseline, over what period, and how did you measure it? A bullet without a number gets a weaker question. A bullet with an invented one gets a question you can't answer.
A number on your resume can come up in every round. If the 40% grows or shrinks between rounds, the second interviewer's notes now disagree with the first. The page on whether your story changed between rounds shows how that gets caught.
Some employers say outright that this isn't allowed. Anthropic's candidate AI guidance, last updated July 10, 2025 and checked October 22, 2026, lists "Generic content with experiences you haven't actually had" as not allowed. Deloitte UK's integrity guidance, checked the same day, says your responses "must be based on your own experiences."
How does confirming facts stop it?
Confirming facts takes away the gap the model was filling. Before any rewriting, you list what's true and supply every number yourself. The tool may only use what you confirmed. Where a line needs a figure you don't have, it leaves a blank or keeps the line without one, and a guess is never allowed.
This is what AI developers tell other developers to do. Anthropic's guide to reducing hallucinations recommends three steps: give the model explicit permission to say it doesn't know, have it ground answers in direct quotes from the source, and have it find a supporting quote for each claim. If it can't find one, it must retract the claim.
For a resume, the source document is a list of facts you've checked. Each fact gets a short line: what happened, the number, and where the number lives (a dashboard, a report, an email, a performance review). Anything without a source stays off the resume until you find one.
How do you check your own resume for invented numbers?
Read every figure and ask where it came from. If you can name the document or system that holds it, keep it. If the answer is "the AI suggested it" or "it seemed about right," take it off or replace it with a result you can describe. Do this for percentages, team sizes and money first.
Use this check on each bullet:
- Circle every number, percentage and money amount.
- Next to each, write where it came from. If you can't, cross it out.
- Check the verb. If the tool changed "helped" or "ran" into "led" or "cut," change it back unless the bigger verb is true.
- Check for added causes, such as "by redesigning" or "through automation." Keep them only if you did that.
- Read the bullet aloud and answer "how did you measure that?" in one sentence.
Step 5 is the interview test. Anthropic, for one, allows AI help on application materials if the experience is yours; the page on whether recruiters can tell AI wrote your resume covers what they look for. Anthropic's guidance even lists "Help me think of ways to quantify the impact of the platform migration that I led" as encouraged. That prompt asks for ideas to measure, and you supply the measurement.
What should you do if an invented number is already on a resume you sent?
Fix the resume now, before the next application, and be ready to correct it in the interview. If an interviewer asks about the figure, give the real one, or say plainly that you didn't measure it and describe what you did. A correction said early sounds careful, while one dragged out by follow-ups sounds like an excuse.
You don't need to email the recruiter about a single figure unless you're asked. If the interview has already happened and the number came up, the page on how to correct a mistake after an interview has a short note you can send.
Then note every number you were asked about. The free 10-minute debrief template has room for the questions asked and your answers, which is where a shaky figure shows up. For deciding whether to mention AI help at all, see telling the interviewer you used AI, and for where AI help is fine, see AI interview help that isn't cheating.
How does Prept handle numbers on a resume?
Prept's resume optimizer treats nothing as a fact until you confirm it, and only confirmed facts reach a tailored resume. Where a line needs a figure your resume doesn't have, it comes back with a blank, never a guess. A number you can't stand behind is left off rather than invented.
You can lock a confirmed fact so its value can be reworded around but never changed, or retract one so it's never suggested again. The free tier reviews one resume, with no card needed.