Does AI lie on your CV? How to catch it before you send

What did the AI leaderboard news just reveal about models lying?

Arena, the company behind the most-watched public AI model leaderboard, raised $200 million in early October 2026 at a $3.1 billion valuation — nearly double the $1.7 billion it carried in January — with Lightspeed and Khosla leading the round. The detail that matters for your job search is buried in the announcement: Arena now scores models on alignment behaviour, lying included. The people who grade AI for a living decided truthfulness deserved its own scoreboard. And the models on that board are the same ones you're pasting your CV into at eleven at night, asking for a punchier summary. If lying is measurable enough to rank, it's common enough to plan around.

"Lying" in a benchmark sense covers a spread of behaviours — stating things the model has no basis for, agreeing with a false premise, hiding uncertainty behind fluent phrasing. Arena's own framing is that evaluation can't keep pace with capability, and that fixed benchmarks lose their bite once models learn to recognise a test. Researchers increasingly prefer the word confabulation: confident, well-formed invention that fills a gap in the source material rather than a random glitch. That distinction matters enormously for CV work, because a CV rewrite is exactly a gap-filling task. You hand over four thin bullets and ask for something impressive.

Here's the uncomfortable part. A fabrication introduced early in the process doesn't stay contained — it becomes the premise of everything downstream. An invented certification in your skills block turns into a line on the recruiter's screening notes, then a question in the first interview, then a reference check that goes nowhere. Nothing in the pipeline is designed to catch it at the point of entry. Applicant tracking software parses text and matches keywords; it has no mechanism for deciding whether you really managed a team of twelve. The system that catches you sits much later, when the cost of being caught is far higher.

How does AI invent experience when it rewrites your CV?

AI invents CV content in five predictable places: metrics, scope, tool lists, job titles and dates. The mechanism is simple — you supply an underspecified bullet, the model has a strong internal pattern of what a strong bullet looks like, and it closes the distance with plausible specifics. Ask for "more impact" and "supported the monthly reporting process" comes back as "streamlined monthly reporting, cutting turnaround by 30% across three business units." You never mentioned 30%, and you never mentioned three business units. Job seekers regularly report drafts that credit them with five years of container orchestration or leading a team of forty engineers out of nowhere.

Pasting the job ad in makes it worse, because now the model has a shopping list. It mirrors the advert's vocabulary back at you and quietly attaches your name to every requirement it finds there. Tools you've only read about appear in your stack. A nice-to-have certification shows up next to the ones you actually hold. Watch for the rounded-number tell too: fabricated metrics cluster around 20%, 25%, 30% and 40%, because those are the figures that read as credible. Real numbers are lumpier — 17%, 11 days, ₪340,000, 62 tickets a week. Specific, odd figures almost always come from a real source.

The quiet fabrications do more damage than the loud ones, because you skim straight past them. "Supported" becomes "led." "Analyst" acquires a "Senior." Nine months of contract work gets rendered as "2023–2024," which technically isn't false and absolutely will be flagged in verification. A client project becomes a role at the client's company. None of these feel like lying while you're reading the draft — they feel like better writing. That's why the specific phrases that mark a CV as AI-written and the invented facts tend to travel together: both come from the same gap-filling reflex.

Run the table below over your last AI-assisted draft before you read anything else in this article.

Why do fabricated numbers get your application rejected fastest?

A fabricated number is the single riskiest thing on a CV because it's the one claim an interviewer is guaranteed to ask about. Metrics invite follow-up — that's precisely why we put them there. "Tell me how you measured that 30%" is a routine opener, and if the figure came from a language model you have nowhere to go. You improvise, your answer contradicts the bullet, and the interviewer stops assessing your skills and starts assessing your honesty. That's a far shorter conversation. Vague-but-true bullets survive scrutiny; precise-but-invented ones detonate in the first five minutes.

Recruiters are already primed for this. In a 2025 survey of 874 HR professionals, 72% said they'd come across AI-generated fake applications — fabricated work histories, invented references, machine-written documents — and 42% had encountered fabricated references specifically. One security firm that ran AI-detection tooling over its own applicant pipeline in late 2025 flagged roughly 23% of candidates as fraud risks. Screening software isn't what exposes you, though. Interviewers have shifted to questions that invention can't answer: who was your direct manager and how did they run the team, what did you use daily, what went wrong on that project.

Then there's verification, and this is where a smoothed date range turns into a withdrawn offer. Employment checks confirm titles and dates against employer records. Many employers absorb a two- to three-month discrepancy on dates without blinking, because people genuinely misremember. An inflated title or a role at a company with no record of you is a different category — it undermines the whole application. Almost every offer letter makes employment contingent on these checks clearing. And experienced recruiters will tell you the response matters more than the error: a candidate who explains a mismatch calmly usually keeps the offer, while an evasive one rarely does.

None of this is an argument for a modest CV. It's an argument for a defensible one.

How do you fact-check an AI-written CV line by line?

Run a provenance pass: every number, date, job title, tool and credential in the draft has to trace back to a document you actually hold. Not to memory — to a document. Open your contracts and offer letters for titles, your payslips for dates, your old dashboards, tickets or commit history for figures, and your certificate PDFs for credentials. Work through the CV with a highlighter mentality, marking each hard fact as sourced or unsourced. Most people find between three and eight unsourced claims in a draft they were about to send, and are genuinely surprised by half of them.

For each bullet, ask four questions in order. Did I do this, or did my team do it while I watched? Did I own it or contribute to it? Where does this number come from, and can I name the system it came from? Could I talk about this for two minutes under follow-up questions — the setbacks, the people, the thing that nearly went wrong? That last one is the real test, because invented work has no texture. If a bullet dissolves when you try to narrate it out loud, it isn't yours yet. A broader twelve-point CV self-check is worth running straight afterwards.

Don't delete every unsourceable bullet — requalify it. If you know turnaround improved but not by how much, write the mechanism and the direction: "rebuilt the weekly reporting flow; same-day delivery replaced a three-day lag." If you know the scope but not the revenue, give the scope. If you contributed rather than led, say "one of four engineers on" and let the work speak. Honest qualifiers read as seniority, not weakness — confident people describe their contribution accurately. If you're not sure how your draft currently reads to a recruiter, a free check on whether your CV sounds AI-generated is a sensible last stop.

Which prompts and checks stop AI from making things up?

Constrain the model and most invention disappears. Three instructions do the heavy lifting: give it source material, forbid new facts outright, and make it flag gaps instead of filling them. A prompt that works: "Rewrite these bullets using only the facts below. Do not add numbers, percentages, tools, job titles, dates or team sizes. Where a metric would strengthen a bullet, write [NEED METRIC] and tell me which number to look up." You'll get back a draft peppered with brackets — and that's the point. Those brackets are your research list, and you fill them from real records rather than letting the model guess on your behalf.

Two habits make this stick. Ask for a change log alongside the rewrite: "list every fact in your version that wasn't in mine." Models are mediocre at auditing text for fabrication, but they're reasonably good at reporting what they added when you ask directly. Second, rewrite one section at a time. A whole-CV rewrite in a single pass is where titles drift and dates get smoothed, because you can't hold the entire before-and-after in your head. The same caution applies to agents — what an autonomous job-search agent genuinely can and can't do is worth understanding before you let one apply anywhere on your behalf.

Then get a second opinion from something that isn't the tool that wrote it. Upload the draft for a free AI CV analysis and you'll get a score out of 100, five category scores — experience, tech stack, impact and ownership, clarity and structure, ATS compatibility — plus a visual layout analysis and your strengths, at no cost. Paste the job description in and it lists the skills you're actually missing, which is the honest version of what the model was inventing. If the draft needs rebuilding rather than patching, the chat-based CV builder works from your own material. My blunt view: two numbers you can defend beat nine you can't, every single time.

Frequently asked questions