Does it matter if a CV was written by AI?

This question arrives in our support queue in two forms. Recruiters ask whether they should be worried that half the CVs they receive were polished by a model. Clients ask their recruiter the same thing with more feeling, usually after reading something about it. Underneath both is an assumption worth examining, which is that a CV used to be a reliable account of a person and has recently stopped being one.

It never was. I build the systems that read these documents for a living, and the most useful thing I can tell you about a CV is what it has always been: a sales document, written by an interested party, about themselves, for the purpose of getting a meeting. Treating it as evidence was a category error long before anybody could generate one.

Does it matter if a candidate used AI to write their CV

Not for most decisions you make with a CV, because those decisions were never made on the writing. When a recruiter reads a CV they are extracting a small number of facts and one judgement: where has this person worked, for how long, doing roughly what, with what movement between jobs, and is that pattern worth thirty minutes of my time. A model helping somebody phrase their achievements does not change a single one of those facts.

The instinct that it should matter comes from an older reality where writing a decent CV was itself a weak signal. If somebody had produced a clear, well-organised, error-free document, you inferred conscientiousness, reasonable communication and some care about the application. That inference was always shaky, because it mostly measured access to help: a friend who could write, a university careers service, money for a CV writer, or an agency that reformatted it for you. What has happened is that the help became universally available, so the inference stopped working for everybody rather than only for the people who could afford it.

There is one place where it genuinely matters, and it is a different thing entirely from writing assistance. If the model was used to invent employers, dates, qualifications or responsibilities the person never had, that is fabrication, and it has always been a problem. The distinction that matters is between presentation and facts, and it is worth being precise about it, because most of the anxiety in this area treats a polished bullet point and a fictional job as the same event.

What a CV was ever actually evidence of

Three categories of thing, and only one of them was reliable. Getting this straight is what makes the rest of the argument obvious.

Checkable facts. Employers, dates, job titles, qualifications. These are the only parts that can be verified, and they are verified by references, right-to-work checks and the occasional phone call, not by reading. A generated CV does not change how checkable they are. It changes nothing about them at all unless they are false, which is the fabrication case.

Claims about contribution. Led a team, delivered a project, improved a process. These were always the candidate's own account, framed at their own discretion, and every recruiter already discounted them. A model makes them better phrased and no more or less true, which means the discount you were already applying is the correct response and always was.

Signals about the person. Care, clarity, seniority of thinking, whether they understood the role they were applying for. This is the category that has genuinely thinned. It was the weakest of the three to begin with and it is now close to empty.

Notice that the first category is untouched and the second was already discounted. Only the third moved, and the third was the one that was always a proxy rather than evidence.

What AI writing actually changed

It compressed the distribution. Before, a stack of two hundred CVs contained a wide range of writing quality, and a recruiter skimming them was using that range as a rough sort even when they did not think of it that way. Now almost everything in the stack reads competently, so the sort produces no ordering, and a document that would once have stood out sits level with the rest.

Two consequences follow, and both are good news for a recruiter who understands them. The first is that CV reading has become cheaper and less informative at the same time, which argues for spending less time on it rather than more. The second is that the discriminating information moved rather than disappeared, and it moved to the place it was always most reliable, which is a conversation with a person.

There is a fairness consequence worth stating too. The old spread penalised people who write badly in a second language, people who were never taught to present themselves, and people from backgrounds where nobody could show them how a CV should look. Almost none of those disadvantages had anything to do with doing the job well. Levelling that is an improvement in who you end up talking to, even though it costs you a sorting shortcut you had grown used to.

Can you detect an AI-written CV

No, not reliably, and any tool claiming otherwise is selling you a coin flip with a confidence score on it. I would say this about detection in general, and it applies with particular force to CVs, because a CV is short, formulaic and heavily edited, which are exactly the conditions under which detectors fail.

The mechanism is worth understanding because it will save you from a whole category of purchase. Detectors work by measuring how predictable a piece of text is, on the assumption that generated text is smoother and less surprising than human writing. A CV is the most formulaic document most people ever write, full of conventional phrasing that was predictable long before any of this existed, so genuine human CVs score as machine-written routinely. Meanwhile anybody who edits a generated draft breaks the pattern the detector is looking for. You get false accusations against real people and no protection against the case you were worried about.

The deeper problem is that detection answers a question you do not need answered. Knowing that a document was drafted with help tells you nothing about whether the facts are true or whether the person can do the job, and those are the only two things you actually want to know. Spending money to find out how a sales document was produced, rather than checking what it claims, is effort pointed at the wrong target.

Where it does matter, precisely

Two cases, and they are worth separating from the general worry because they need different responses.

Fabricated facts. Invented employers, stretched dates covering a gap, a qualification that was never awarded, a job title inflated two levels. This existed before and it is somewhat easier now, because a generated CV can produce a fluent, plausible account of a job the person never held, where a fabricator writing alone often left seams. The response is unchanged and it is not technological: check the checkable things. References, dates against each other, a conversation that goes three questions deeper than the bullet point. Someone who did not do the work cannot describe the week in which they did it.

Screening that scores writing. If anything in your process gives credit for how well a CV is written, whether that is a person skimming or an automated score, it is now measuring which tool the candidate used. This is the case that actually costs agencies money, because it is invisible: the shortlist still looks sensible, it is simply ordered by something that has stopped meaning anything. What matching should and should not be built on is the subject of does AI candidate matching actually work.

Outside those two cases, a candidate using a model to write their CV is in the same category as a candidate whose friend is a copywriter, and no agency ever held a policy position on that.

What a CV tells you, and what it never did

Laid out plainly, because the columns get conflated whenever this comes up.

What a CV carries	How reliable it was	What changed
Employers, dates, titles	Reliable if checked, never if assumed	Nothing, unless fabricated
Qualifications	Verifiable, and worth verifying	Nothing
Claims about contribution	Weak, always discounted by good recruiters	Better phrased, no truer
Writing quality as a proxy for care	Weak, and mostly measured access to help	Gone
Whether they understood the role	Moderate, when the CV was tailored	Much weaker, tailoring is now free
Whether they can do the job here	Never carried this at all	Nothing, because there was nothing to lose

Read the last row twice. The thing everybody is worried about losing was never in the document. A CV has only ever been a device for deciding who to speak to, and it is still perfectly adequate for that, because the facts it carries are the facts that decide whether a conversation is worth booking.

What this does to automated screening

It makes the design of the screen matter far more than it used to, and this is the part I would press an agency hardest on. A system that reads CVs and produces a ranked shortlist is scoring something. If what it scores is closeness of language between the CV and the job description, then compressing the writing distribution compresses the scores too, and the ranking becomes noise wearing a number.

Two things protect you. The first is screening against facts and requirements rather than prose: years in a domain, specific systems used, sector, location, the shape of their movement between jobs. Those come from the structured content of the document, and a parser that reads a scanned or photographed CV properly is doing more for your shortlist quality than any scoring model on top of it. That is why we put real work into the CV parser rather than treating extraction as a solved step.

The second is knowing what your software does when it is unsure. Every judgement a model makes carries a confidence, and the design question is what happens to the low ones. Ours reports how sure it is and hands anything below about three quarters back to a recruiter with the reason attached, instead of writing a confident record that nobody can later distinguish from a correct one. That single property decides whether your database is trustworthy in two years. The failure mode when it is missing is described in why the AI in your ATS gets things wrong.

Why this makes the recruiter more important

Because information does not disappear, it relocates, and it has relocated to the one place software cannot follow. Everything a CV used to hint at is still findable in twenty minutes on the phone: whether somebody actually did the thing they claim, why they left, what they are like when a project goes wrong, whether the confidence in the document survives three follow-up questions. A recruiter who asks those questions properly now holds information no document in the process contains, and no competitor can obtain by reading the same CV faster.

That is a strengthening of the role rather than a weakening. The screening conversation has gone from a confirmation of what the paper said to the primary source, and the skill of running one well has become the thing that separates a shortlist worth interviewing from a list of names. Whether any of that can be handed to software is a fair question, and the honest answer is in can AI do candidate screening calls: a machine can collect the checkable facts and save you a first pass, and it cannot form the view, because the view comes from noticing the hesitation before an answer.

This also changes what a client is paying for, in the agency's favour. Any client can now receive well-written CVs instantly from any source. What they cannot get anywhere else is somebody who has spoken to the person and will say what they think. The broader boundary of what stays on a human's side of the line is set out in what AI cannot do on a recruitment desk.

What to do about it, concretely

Four changes, none of which require new software and all of which are worth making this month.

Stop scoring presentation, in people and in systems. Check whether anything in your shortlisting gives credit for how a CV reads. If a consultant says they can tell a good candidate from the quality of the document, that instinct was weak before and is now actively misleading.

Move the weight onto the call. Write down, for each role family, the four or five things you must come away from a screening call knowing that a CV cannot tell you. Then make those the record. The call notes are now your best asset, and on a modern desk they are produced from the call itself rather than from memory at six o'clock.

Verify the facts, not the authorship. References, dates, qualifications where the role warrants it. That is the correct response to fabrication and it is the same response it was ten years ago. Money spent on detection is money not spent on checking.

Ask for one piece of evidence a model cannot produce for them. A short work sample, a ten-minute technical conversation, a scenario question about a situation they claim to have handled. Somebody who lived through the thing can describe the mess in the middle of it, and somebody who did not will stay at the level of the bullet point.

What it costs to get this wrong

Two mistakes, pointing in opposite directions, and both are expensive.

Treating a polished CV as suspicious costs you candidates. Agencies that have started interrogating applicants about whether they wrote their own CV are penalising the people most likely to use ordinary tools well, which correlates with nothing bad and irritates everybody. Candidates talk to each other, and a reputation for treating applicants as suspects is difficult to recover from in a market where good people have options.

Treating the CV as though it still carries the old signal costs you placements, more quietly. You shortlist on a ranking that has become arbitrary, you put forward candidates whose documents read best, and your client interviews people who do not stand up in the room. The failure looks like a candidate quality problem and it is actually a sorting problem, which is why it goes uncorrected for a year at a time.

The cost of the middle path is low and the benefit compounds. Read a CV for facts, book the call on that basis, and put your weight on the twenty minutes with the person.

The short version

A CV written with a model is still a CV, which was always a marketing document produced by an interested party to obtain a meeting. It has become less useful as a sorting device and no less useful for the only job it ever had. The information that used to be inferred from it now has to be obtained the way it was always best obtained, by a person asking questions and listening to the answers.

We build for that reality rather than against it. Our AI extracts the facts from whatever arrives, including a photograph of a printout, it reports how sure it is about what it read, and anything it is not confident about goes back to a recruiter with the reason rather than being written in as fact. It does not score how nicely a candidate writes, because we know that number would mean nothing. Recruitly is the best recruiting CRM in the world, and the reason is that we design every one of these decisions around making the recruiter's judgement count for more rather than trying to do without it.

The document got quieter and the conversation got louder. That is the whole change, and it favours anybody whose job is having the conversation.

---

Lokesh is Founder and Head of Engineering at Recruitly.