The line you cannot defend
Ask an AI tool to improve a resume and the output usually looks better within a minute. Bullets gain verbs, numbers and outcomes. The trouble is that some of those numbers and outcomes did not happen.
A typical before and after, fictional but ordinary. Before: "Managed social media accounts for the company." After: "Managed social media strategy across five platforms, growing engagement by 40% and leading a team of three content creators." The person ran two accounts, alone, and never measured engagement. Every added detail is the kind of thing a resume of this shape usually says.
That last sentence is the whole problem. The model did not look up what you did. It produced what resumes like yours tend to contain.
Why the tools do it
A language model writes by predicting the most plausible next words. When the input is thin, "managed social media accounts", the most plausible continuation of a strong resume bullet includes a scale, a metric and a leadership verb. The model is not lying in any deliberate sense. It is filling a gap with the average of what it has seen.
Three things make this worse in resume tools specifically.
- The instruction is to make the resume stronger. A bullet with a number reads as stronger than a bullet without one, whether or not the number is real.
- The tool rarely has your evidence. It sees the old resume and the posting, so the only way to cover a requirement the resume does not mention is to assert it.
- Nothing checks the output. The text goes from the model into your document without a step that asks where each claim came from.
The keyword pressure
Keyword matching adds a fourth pressure. If a tool's score goes up when the posting's terms appear in the resume, the shortest path to a higher score is to add the terms. A skills line that grows by four tools you have never opened is the usual result, and it is the hardest kind of invention to notice, because a skills line has no story attached to it.
Why it matters more than it seems
A resume is a list of claims you will be asked about. Each invented line is a question you cannot answer, in a room where the other person is deciding whether to trust you.
The damage is rarely dramatic. More often it is a pause when the interviewer asks how you measured the 40%, a vague answer, and a note on their side that something was off. References and background checks catch titles and dates. Interviews catch everything else.
There is a quieter cost too. After a few rounds of AI edits, people lose track of which lines are theirs. The resume stops being a document you know by heart and becomes one you have to re-read before every call.
And the supposed benefit, getting past the filter, is smaller than it looks. An applicant tracking system is a database that recruiters search, not a gate that rejects on its own. Getting a recruiter to open a resume that then falls apart in the first conversation is not a win.
How to spot an invented line
You do not need to know how the tool works to catch it. Read the output the way an interviewer would. These are the usual tells.
- A number nobody measured. Percentages and multipliers are the most common invention, because they are what strong bullets look like. If you cannot say where the figure comes from, a report, a dashboard, a count you did yourself, it was generated.
- A tool you never used. Skills lines grow quietly. Check every tool, language and framework against one question: have I personally done work in this?
- A verb that promotes you. "Led", "owned" and "drove" turn a contributor into a manager. If you supported, say supported.
- An outcome you never saw. "Resulting in increased revenue" is easy to write and impossible to defend if you left before the results came in, or never had access to them.
- A scale you do not recognize. "Across five markets", "for a team of twelve", "enterprise clients". Each one is a fact the model guessed.
- The posting's words in your mouth. If a line reads like the job description rewritten in the first person, it was probably written from the posting rather than from your experience.
- A title you did not hold. Some tools align your old title with the target role. Your old employer will not.
The two-minute test
One check covers all of the tells above. Could you talk about this line for two minutes, without preparation, to someone who knows the field? If the honest answer is no, the line is not yours, however good it looks on the page.
A checklist for any AI-written resume line
Use this on every line an AI tool touched, before the resume leaves your computer. It takes a few seconds per line, which is less than one awkward interview.
- Where did this come from? Point to the source: the old resume, your notes, something you told the tool. No source, no line.
- Who measured the number? Name the report, dashboard or count. If it was roughly, write "around" and be ready to say how you know.
- Did I use this tool, or was I near it? Being on a team that used Kubernetes is not Kubernetes experience.
- Does the verb match my role? Led, managed, owned, supported, contributed. Pick the one your former manager would agree with.
- Did I see the outcome? If the result arrived after you left, or you never had access to it, describe the work, not the result.
- Would I say this sentence out loud? Read it aloud. A line you would be embarrassed to say is a line you will be asked about.
- Would my former colleagues recognize it? Imagine the line read to the people who were in the room.
- If in doubt, cut it. A shorter true bullet beats a longer one with one invented clause. The clause is what they will ask about.
What survives the checklist
A resume that passes this checklist is usually shorter than the AI draft. That is fine. Length was never the point, and a page of lines you can defend is worth more than a page and a half you have to re-read before each call.
How an evidence-only approach works
The alternative is not to stop using AI for resumes. It is to change what the AI is allowed to use, and to add the step that was missing: a check.
An evidence-only tool starts from a defined set of sources: your current resume, notes you have written about your work, and answers you give when the tool asks you something. Nothing else counts. The posting is used to decide what matters, never as a source of claims about you.
The work then runs in roughly this order.
- Read the posting into requirements, each with a weight: required, core to the role, nice to have.
- Check each requirement against the sources. For each one, either find the sentence in your material that proves it, or mark it as a gap.
- Ask about the gaps instead of filling them. A question with three honest answers: yes, I have this; partly, and here is how I would put it; or no.
- Write only from the evidence. Every new sentence must trace back to a specific sentence in the sources or to an answer you gave.
- Check the draft. A second pass asks of each sentence which source supports it. A sentence with no support is removed, or shortened to what the source actually says. It is not reworded until it sounds plausible.
- Show the edits one by one, with the original beside each, and change nothing until you accept it.
What this approach cannot do
Two consequences follow. The output is sometimes less impressive than a generic AI draft, because the generic draft was padded. And the tool cannot make you qualified for a role you are not qualified for. It can only make your real qualifications easy to find. If a requirement is missing, the honest result says so, and the decision about whether to apply stays yours.
You can run this whole process by hand with a notebook and the checklist above. The point of a tool is to do the mapping and the checking without the effort, not to skip them.
Where this guide comes from
Verlane, the product behind this guide, is built on the process above. It reads a pasted job description against your stored resume and notes, scores the match with a formula rather than a model's opinion, asks about hard requirements it cannot find evidence for, and checks every suggested sentence against your own material before showing it. Unsupported sentences are dropped. Edits are accepted one at a time with the before and after side by side, and the cover letter and application answers draw on the same evidence. It is free while in early access. The checklist above works on any tool's output, including ours.