Career & Jobs
Resume Writing When Both Sides Use AI
By Jim Vernon, Editor, AI Intelligence International · Published 28 January 2026 · Reviewed against our editorial standards · About the author
Application volume has risen sharply because generating an application costs nothing. Screening has become model-assisted in response. The result is an arms race in which both sides produce and process more text, and nobody reads more carefully.
In that environment, the winning CV is not the most polished. It is the most specific, because specificity is the one property that generated text reliably lacks.
Key takeaways
- What screening models actually reward: Most automated screens are looking for evidence of required skills, tenure patterns, and role-relevant nouns.
- The generated-CV tell: Experienced recruiters have learned the signature of machine-written applications: uniformly balanced bullet lengths, three-part parallel structure everywhere, abstract impact verbs, and an absence of anything awkward or specific.
- Using AI to draft without sounding generated: Use the model for structure and compression, not for content.
- Cover letters still matter, but differently: A generic cover letter is now worse than none, because it signals that you spent nothing on the application.
What screening models actually reward
Most automated screens are looking for evidence of required skills, tenure patterns, and role-relevant nouns. They are pattern matchers, and they are reasonably good at recognising that 'led migration of billing system' matches a requirement for systems migration experience.
What they are poor at is inferring seniority from vague achievement language. 'Drove significant improvements across key initiatives' contains no matchable content. It survived the human era because a person could read tone; it fails a keyword-and-evidence screen.
Write in nouns and numbers. Systems, tools, regulations, client types, volumes, budgets, team sizes. Each one is a hook a screen can catch and a human can ask about.
The generated-CV tell
Experienced recruiters have learned the signature of machine-written applications: uniformly balanced bullet lengths, three-part parallel structure everywhere, abstract impact verbs, and an absence of anything awkward or specific.
The fix is not to write badly. It is to include details that no model could have invented about you: the name of the internal system, the odd constraint you worked around, the client sector, the thing that went wrong.
One concrete sentence about a failure and what you changed afterwards will do more for credibility than a page of achievements, precisely because generated text almost never volunteers a failure.
Using AI to draft without sounding generated
Use the model for structure and compression, not for content. Give it your raw notes — messy, specific, over-long — and ask it to tighten each bullet to one line while preserving every number and proper noun. That inverts the usual failure mode, where the model supplies the language and you supply nothing.
Then rewrite the top third by hand. The summary and first role are what a human actually reads, and they should sound like you on a good day, not like a well-behaved assistant.
Finally, ask the model to critique rather than produce: 'which of these bullets could apply to anyone in this profession?' The ones it flags are the ones to cut.
Cover letters still matter, but differently
A generic cover letter is now worse than none, because it signals that you spent nothing on the application. A short letter that references something specific — a product decision, a recent announcement, a problem visible from outside — signals attention that cannot be mass-produced.
Three paragraphs is plenty: why this organisation specifically, what you would expect the first ninety days to involve, and one piece of evidence that you have done the hard part before.
If you cannot write the first paragraph without research, that is the point of the exercise.
The portfolio problem
In fields where output can be generated, the portfolio's role has shifted from proving you can produce to proving you can judge. Annotated work — here is what I chose, here is what I rejected, here is what I would change now — is far more persuasive than finished artefacts.
This is also more robust to the obvious suspicion. Anyone can present a beautiful deliverable; explaining the trade-offs behind it in your own voice is the part that survives scrutiny in an interview.
Keep it small. Three annotated pieces beat a gallery, because a gallery will not be read.
Running the checks
Score your CV for specificity and role fit before sending it, and pay attention to the bullets flagged as generic rather than the overall number.
Then align your public profile: the headline tool is useful mainly because it forces you to state what you own rather than what you are, which is the same discipline that makes a CV work.
Two readers, one document
Your CV is now parsed by software before a person sees it, but the person is still the one who decides. Optimising only for the parser produces keyword soup that a human skims and discards; optimising only for the human risks never reaching them.
The document that satisfies both is plain and specific: standard section headings, a single column, no text inside images or tables, and achievements written with numbers. Parsers handle it cleanly and humans read it quickly.
Skip the hidden-keyword tricks — white text, invisible blocks, stuffed footers. Modern screening flags them, and a recruiter who spots one stops reading immediately.
Rewriting one bullet, three ways
Weak: 'Responsible for customer support.' It states a duty, contains no evidence, and matches thousands of other applications.
Better: 'Handled first-line support for a SaaS product.' Now it says what and where, but there is still nothing to distinguish the person who did it well from the person who did it badly.
Strong: 'Handled first-line support for a SaaS product, ~60 tickets/day, cut median first response from 4h to 40m by rebuilding the macro library.' Specific, checkable, and it naturally contains the terms a parser is looking for without any keyword stuffing.
Frequently asked questions
Do applicant tracking systems reject AI-written CVs?
They do not detect authorship reliably. What they reject is a CV without matchable evidence, which generated CVs often are because they favour abstraction.
How long should a CV be now?
Two pages for most experienced roles. Volume no longer signals effort, so extra length only dilutes the specific details that differentiate you.
Should I include an AI skills section?
Only with specifics: which tools, in which workflow, with what result. 'Familiar with AI tools' is now equivalent to listing email as a skill.
Is it worth applying to hundreds of roles?
No. Response rates on mass applications have collapsed precisely because everyone can send them. Five researched applications a week outperform fifty generated ones.
Can I use AI to write my CV?
To draft and tighten, yes. Every claim must still be true and yours — invented specifics fall apart in the first interview.
Should the CV be tailored per application?
Adjust the summary and the top three bullets to match the posting. Rewriting the whole document for every role is rarely worth the hours.