Almost every student now runs their resume through a chatbot before sending it. That is sensible: rewriting a flat bullet into a sharp one is exactly the kind of task these models are good at, and doing it by hand is slow.
It also fails in ways that are easy to miss and hard to recover from. We run a language model over resumes here, and the rules we had to write to keep it useful are a decent map of where the failures are.
1. It invents numbers
This is the serious one. Ask a model to strengthen a bullet and it will often hand back something like:
Optimized the data pipeline, improving processing speed by 40% and reducing costs by $12,000 annually.
You never said 40%. You never said $12,000. The model produced plausible-looking figures because resume bullets usually contain them, and it is completing a pattern.
You will not notice, because it reads exactly like what a good bullet looks like. The interviewer will notice, when they ask how you measured it.
Our own analyzer is explicitly forbidden from doing this. When a metric would help but is not in the resume, it has to write a placeholder like [X%] or [N users] rather than a number — an obvious blank you have to fill in yourself, instead of a confident fabrication you have to catch.
What to do: after any AI rewrite, read it line by line and delete every number you cannot personally source. If a bullet reads worse without the invented figure, that tells you the underlying work needs describing better, not decorating.
2. It appends outcomes you did not achieve
Subtler, and more common. The model reaches for closing phrases that sound like impact:
Built an internal dashboard, enabling faster decision-making across teams. Migrated the service to containers, ensuring improved reliability.
Did decisions get faster? Did reliability improve? You do not know. Nobody measured. These clauses are filler wearing the costume of a result, and experienced recruiters read straight past them, so you spent a line and gained nothing.
What to do: cut any clause starting enabling, ensuring, allowing, resulting in unless you can say what actually changed.
3. It gives international students advice that can hurt them
Ask a general-purpose chatbot how to handle work authorization on a resume and there is a real chance it suggests stating your visa status — that you are on F-1, that you will need sponsorship, that you have STEM OPT eligibility.
That is the opposite of standard guidance. Work authorization belongs on the application form, where there is a field for it and a legal process around it. On the resume it is an unprompted disclosure that gives a screener a reason to stop reading before they have seen your work.
Our analyzer is prohibited from raising the subject at all. A general chatbot has no idea you are on F-1 unless you tell it, and once you do, it will helpfully advise you on something where helpfulness is the wrong instinct.
What to do: keep status off the resume. Answer it honestly on the form. If you want to know which employers sponsor before you apply, that is what the employer directory is for.
4. It writes in a register that reads as generated
Models reach for the same vocabulary: spearheaded, leveraged, synergized, robust, cutting-edge, seamless. A resume where every bullet is polished to the same gloss reads as machine-written, and recruiters have been seeing that gloss for two years now.
The irony is that models also strip the specific, slightly awkward details that make you sound real — the odd framework, the unglamorous constraint, the thing that broke.
What to do: use AI to fix a bullet you have already written, not to write it. Feed it your clumsy version. A clumsy sentence about real work beats a polished sentence about nothing.
What it is genuinely good for
None of this means avoid it. It is very good at:
- Tightening a bullet you wrote that runs too long
- Suggesting stronger opening verbs for lines that start weakly
- Spotting which keywords from a job description your resume does not mention
- Reformatting inconsistent dates and headings
Notice these are all editing tasks on content you supplied. That is the line. The moment it is generating claims rather than sharpening yours, you are no longer the author of your own resume, and you will find out in the interview.
If you want the editing without the invention, the analyzer here runs under the rules described above: it will not add a number you did not give it, and it will not raise your visa status.