Treat AI as an editor, never as an author. Supply every fact — the numbers, the scope, the tools, the outcomes — and use the model to tighten, restructure, and surface keywords. The moment it starts generating achievements, it starts generating things you cannot defend.
The reason unedited AI resumes underperform is not that recruiters run detection software. It is that model output defaults to safe, balanced, slightly grand phrasing, and every candidate using the same tool arrives sounding the same. Generic loses regardless of who or what wrote it.
Below: what AI genuinely does well and badly, three prompts that produce usable output, the specific tells experienced readers recognize, before-and-after edits of real model output, what happens to your data when you paste a resume into a chatbot, and what employers are doing with AI on their side.
Key Takeaways
- Four tasks AI does well and four it reliably ruins
- Three prompts written to produce editable output
- The tells: what unedited output looks like to a reader
- What to strip before pasting your resume anywhere
- Your rights where employers use automated screening
What it does well, and what it ruins
Start Here
The dividing line is whether the task needs facts you own. Editing your material is a strength; producing material is where it goes wrong.
| Task | Verdict | Why |
|---|---|---|
| Tightening a wordy bullet | Good | Pure editing on your facts |
| Listing likely keywords from a posting | Good | Extraction, then you filter |
| Fixing grammar and idiom | Good | Strong for non-native writers |
| Suggesting stronger verbs | Good | You keep or discard each one |
| Writing achievements from a job title | Bad | Produces plausible fiction |
| Inventing metrics | Bad | Numbers you cannot defend |
| Writing your summary unprompted | Bad | Defaults to interchangeable phrasing |
| Judging your career strategy | Bad | No context on your market or goals |
The fabrication risk is the one to take seriously. Models produce confident-sounding tools, certifications, and percentage improvements that were never real, and they do it most readily when your input was thin. If you cannot defend a line for ninety seconds under questioning, delete it regardless of how polished it looks.
Three prompts that produce usable output
Do This Next
Every good prompt here supplies the facts and constrains the output. Vague prompts return vague resumes, which is most of why people conclude the tools do not work.
1. Tightening a bullet. "Rewrite this bullet in under 25 words. Keep every number and tool name exactly as written. Do not add any claim I have not made. Give me three variants: [paste bullet]."
2. Finding keyword gaps. "Here is a job posting and my resume. List the skills, tools, and credentials named in the posting that do not appear in my resume. Do not suggest wording — just the list, grouped by whether the posting calls them required or preferred."
3. Interrogating your own material. "Read this role description from my resume and ask me eight questions that would help me add specific numbers. Do not write anything for me — only ask." This is the most underused prompt of the three, because it turns the model into the interviewer who extracts your metrics rather than the writer who invents them.
Notice the constraints in all three: word limits, do-not-add instructions, and multiple variants so you choose rather than accept. Removing those constraints is what produces the output everyone recognizes.
Two prompts to avoid, because they invite fabrication directly. "Write me a resume for a marketing manager" gives the model nothing but a job title, so it fills the gap with invented achievements that sound like the average of every marketing resume ever written. "Make this sound more impressive" is worse, because it explicitly asks for inflation and the model will supply it — usually by adding scale or outcomes you never claimed.
A third pattern worth knowing: asking for a cover letter from your resume alone. The result is a competent restatement of things the employer can already read, which is exactly what makes cover letters skippable. If you use AI here, give it the specific reason you want this company and the one gap you need to explain — the parts a model cannot infer are the only parts worth including.
The tells experienced readers notice
Watch Out
Nobody is running detection software on your resume. They are noticing that it contains nothing only you could have written.
- Recycled verbs across every bullet. Orchestrated, orchestrated, championed, drove — appearing four times in one role.
- Balanced three-item phrases everywhere. "Strategy, execution, and delivery" in one line and "planning, alignment, and growth" in the next.
- Achievement shapes with no real numbers. "Significantly improved efficiency" and "drove substantial growth" — the grammar of a metric with none present.
- Uniform register. Every bullet the same length and temperature, which no human career actually produces.
- Slightly grand vocabulary in an ordinary role. Nobody "orchestrated" a weekly stock count.
- Nothing specific to your industry. The bullets would fit three unrelated jobs without editing, which is precisely the problem.
The fix for all six is the same and it is not stylistic. Add the details only you know — the system name, the client type, the number of sites, the thing that went wrong and what you did. Specificity is what no model can supply, which is exactly why it is what convinces.
Before and after: editing model output
Quick Win
In each pair the "before" is typical unedited output. The edit adds facts the candidate already had and removes vocabulary nobody uses out loud.
Operations role
Before: Orchestrated process optimization initiatives that significantly enhanced operational efficiency across multiple business units.
After: Rebuilt the stockroom count process across 3 sites; cut shrink from 2.1% to 1.3% and removed 6 hours of weekly manual reconciliation.
Marketing role
Before: Harnessed data-driven strategies to drive substantial growth in customer acquisition and engagement metrics.
After: Ran paid social and search on a $1.2M annual budget; rebuilt attribution in Google Analytics 4 and cut blended acquisition cost 28% over two quarters.
Engineering role
Before: Championed the adoption of modern development practices, fostering a culture of technical excellence within the team.
After: Introduced on-call runbooks and blameless postmortems for a 9-engineer team; production incidents fell 35% across two quarters and mean resolution time went from 74 minutes to 21.
None of the "after" versions is better written in a literary sense. They are more specific, and specificity is the entire difference between a bullet that gets remembered and one that gets skimmed.
Edited it with AI? Check it still parses. Reformatting through a chat tool frequently loses structure. See what a system extracts from the final file — free, no signup.
Run a free ATS resume check →What happens to your data when you paste
Pro Tip
A resume is one of the densest packets of personal data you own. Strip it before pasting, and check the tool's policy once rather than never.
Before pasting into any consumer chat tool, remove your home address, phone number, and email. None of them helps the model rewrite a bullet, and all of them are exactly the fields you would not want retained. Replace employer names with generic descriptors if you are searching confidentially — "a mid-size logistics company" produces the same quality of edit as the real name.
Client names under confidentiality are a genuine issue rather than a theoretical one, particularly for consultants, agency staff, and anyone working under an NDA. Describing the client by sector and size keeps you inside your obligations and loses nothing in the rewrite.
Check whether the tool uses inputs for training. Several consumer products do by default and offer a setting to opt out, while business tiers typically exclude your data. Under GDPR in the EU and UK, and laws such as the CCPA in California, you generally have rights to access and deletion — worth knowing if you have pasted more than you intended.
AI on the employer's side
Reality Check
Employers use it for ranking and matching, not for judging your prose. The parsing rules that decide whether you are readable have not changed.
Workday, Greenhouse, Lever, iCIMS, and Taleo all offer matching or ranking features to some degree, and larger employers use them to order a long applicant list. What sits underneath is the same extraction as always — your file becomes flat text, your experience is inferred from date ranges, and graphics contribute nothing. No amount of AI on either side changes the fact that a two-column layout can arrive as interleaved fragments.
On the regulatory side, several US jurisdictions now require disclosure, candidate notice, or independent bias auditing where automated tools materially inform hiring decisions, and comparable rules are expanding in the EU. You are entitled to ask what is being used and how — reputable employers answer, and the question reads as informed rather than awkward.
The practical consequence for you is unchanged: make the file extract cleanly, put honest terms in it, and stop worrying about outsmarting a ranking model you cannot see.
One category does deserve attention: one-way recorded video interviews scored by software. These remove the interviewer entirely, so structure carries everything — give the direct answer in your first sentence, keep within about three-quarters of the time limit, and look at the lens rather than your own image. Where an automated system evaluates the recording, ask what is being used; in several jurisdictions you are entitled to be told.
It is also worth being sceptical of AI-detection claims in either direction. Tools that purport to identify machine-written text are unreliable and produce false positives on perfectly ordinary human writing, particularly from non-native English speakers. No serious employer is rejecting candidates on the output of one, and you should not write defensively to try to defeat one.
A workflow that keeps the output yours
Key Takeaway
Facts first, model second, your edit last. Reversing that order is what produces the resume everyone recognizes.
- Write the raw facts yourself, badly. Plain sentences, no polish, every number you can remember. This is the step people skip, and it is the only one that cannot be delegated.
- Use the interrogation prompt. Ask the model for eight questions that would help you add specifics, then answer them. This surfaces material rather than inventing it.
- Ask for three variants of each bullet, under 25 words. Variants force a choice; a single suggestion invites acceptance.
- Reject anything you did not supply. If a number, tool, or claim appeared that you did not write, delete it without negotiating with yourself about whether it is roughly true.
- Read the result aloud. Anything you would not say in a screening call goes. This catches the grand vocabulary faster than reading silently does.
- Re-check the file parses. Copying between a chat window and a document frequently strips or mangles structure.
Step four is where discipline actually gets tested. Model output is fluent and it is genuinely tempting to keep a line that sounds impressive and is only approximately accurate. That line is the one an interviewer asks about, because it is the most interesting claim on the page.
A useful habit while drafting: mark every AI-suggested line with a symbol and delete the symbol only once you have verified the underlying fact. Whatever still carries a symbol at the end is what you have not actually confirmed, and it is usually one or two lines more than you expected.
Bottom line
Final Word
Supply the facts, constrain the prompt, edit hard, and delete anything you could not defend for ninety seconds. Then check the file still parses.
The most useful thing these tools do is not writing at all — it is asking you questions that surface numbers you had forgotten. Used that way, the output is more specific than what you would have written alone. Used as a first draft generator, it produces the most forgettable resume you have ever sent.
Try the interrogation prompt on your current role: ask for eight questions that would help you add specific numbers, and answer them honestly. That exercise takes twenty minutes and produces material no model could have invented.
Edited with AI? Verify the final file
Reformatting through a chat tool frequently loses structure that the original had. See exactly what an applicant tracking system extracts — free, no signup, no scan limit.
Run the Free ATS Resume CheckAlso useful: finding your numbers · free resume builder
Frequently asked questions
Often, and not through detection tools — through pattern. Unedited output has recognizable habits: recycled verbs, balanced three-item phrases, achievements with no specific numbers, and a uniform register across every bullet. What gives it away is the absence of the particular details only you would know.
Reformatting, tightening wordy bullets, suggesting stronger verbs, listing likely keywords from a job description, and catching grammar issues. All of these are editing tasks on material you supply. It is a capable editor of your facts and a poor author of them.
Generate achievements, invent numbers, or write your summary from scratch. Models produce plausible-sounding metrics that were never real, and a fabricated figure you cannot defend costs the interview. Supply every fact yourself and use AI only to improve how it reads.
Read the tool's data policy first. Your resume contains your address, phone number, employer history, and sometimes clients under confidentiality. Some consumer tools use inputs to improve models by default. Remove contact details and client names before pasting, and prefer settings that exclude your data from training.
Unedited output does, because it reads generic and generic loses. Heavily edited output does not, because at that point the thinking and the specifics are yours. The risk is not the tool — it is submitting a first draft that says nothing only you could say.
Some do, through ranking and matching features inside their applicant tracking system. Several US jurisdictions now require disclosure, consent, or bias auditing where automated tools materially inform hiring decisions, and similar rules are expanding. You are entitled to ask what is being used and how.
It can surface terms you genuinely have but forgot to write down, which is legitimately useful. It will also over-stuff if left unsupervised, recommending terms you cannot defend. Use it to generate the candidate list, then apply your own judgment about what is true.
Yes, and this is one of its strongest use cases. It fixes idiom, article usage, and register issues that would otherwise distract a reader from your actual experience. Keep your own facts and structure, and use it to make the prose read naturally in the destination market.
Done for you
Turn this advice into an interview-ready resume
Professional writers rebuild your resume for ATS + recruiters — unlimited revisions, interview guarantee.