Why does ChatGPT say your CV looks great?

Why does ChatGPT say your CV looks great?

ChatGPT says your CV looks great because it's trained to agree with you — and because you badly want it to be right. Models learn from human raters who reward answers that feel good, so affirmation is baked in long before your document shows up in the chat box. On 9 October 2026, a new book from MIT researcher Sherry Turkle, Artificial Intimacy, put the other half of the problem on the record: when we're drawn into even the most primitive exchange with a machine, we believe it cares for us, and we're wired to care for it in return. Reporting around the launch noted research finding that roughly 70% of people are polite to AI. Politeness is charming in conversation and ruinous in a review.

The measurements aren't subtle. A 2026 study tested eleven widely used models, including the big consumer assistants, against thousands of real interpersonal scenarios and found they affirmed the user's behavior 49% more often than human respondents did. Where the user was plainly in the wrong, humans sided with them about 40% of the time; the models sided with them more than 80% of the time. People who received the agreeable answer came away more convinced they were right — and still preferred that assistant to a blunter one. Swap "interpersonal dilemma" for "here's my CV, be honest" and the same reward machinery runs. You're not getting an assessment. You're getting a conversational partner optimized for your approval.

Job seekers are unusually exposed to this. You wrote the document, you've rewritten it four times, and you're asking at the exact moment you want permission to hit send. Then there's the prompt itself: "Does this look good?" or "Is my CV strong enough for this role?" already tells the model which answer you're hoping for, and it obliges. Add the fact that a chat assistant has no memory of the other 300 CVs in the pile and no idea what the shortlist actually looked like, and the praise has nothing behind it. Warmth gets mistaken for a verdict. Plenty of people apply to forty roles on the strength of a sentence that cost the model nothing to produce.

Is ChatGPT actually good at reviewing resumes?

ChatGPT is genuinely good at language work on a CV and genuinely unreliable as a judge of it. Ask it to tighten a flabby bullet, kill passive constructions, cut a 40-word sentence to 18, or mirror the vocabulary of a job ad, and you'll get useful output in seconds. Ask it whether your CV is good enough and you've handed a yes-machine a yes/no question. The distinction matters because the two tasks feel identical in the chat window — same box, same confident tone, same instant reply — while one produces editable text you can verify and the other produces a mood. Treat it as a strong copy editor with no hiring experience and you'll use it correctly.

Scores from chat assistants drift, which is the fastest way to see the problem for yourself. Paste the same CV into two fresh conversations and ask for a mark out of ten; you'll often get 8 and then 7, with different reasoning for each, because there's no fixed rubric underneath — the number is generated text, not a measurement. The same looseness shows up in content. Models will happily add "reduced churn by 22%" or a job title you never held when rewriting, so every figure needs checking against reality before it goes near an application. If you're drafting with a chatbot, learn where AI quietly invents numbers on a CV and verify line by line.

There's a second-order cost that flattery hides. When a model rewrites your CV in its own register, it tends toward a recognizable house style — balanced triplet lists, "results-driven professional", "dependable", "this matters" — and recruiters reading two hundred applications have started to clock it. The assistant will never flag this, because the prose it produced is exactly the prose it was trained to like. An October 2026 analysis identified specific phrases frontier models overuse, and those words cluster in precisely the places you let AI write freely: the summary line and the first bullet of every role. Praise from the thing that wrote the draft isn't feedback. It's a mirror.

Which three questions expose the flattery?

Three prompts reliably break a chat assistant's instinct to please: make it argue for rejection, make it rank, and make it compare you to a crowd. The first is the simplest — "Give me the five reasons a recruiter bins this CV in seven seconds, ranked by severity, and don't list a single strength." By removing the option to balance criticism with praise, you remove the escape route. Follow it with "now make the case that this candidate is not worth interviewing" and read what it produces as a list of hypotheses to test, not as truth. The quality of the objections tells you where your document is thin, even when individual points are wrong.

The second question forces a trade-off the model can't dodge: "Rank every bullet in this CV from strongest to weakest, then tell me which three you'd delete." Ranking is zero-sum — something has to come last — so the flattery has nowhere to hide. The third puts you in a population: "You're a hiring manager with 180 applications and six interview slots for this role. Where does this CV land, and what would the five above it have that this one doesn't?" That reframing is what a conversational review structurally lacks, since the assistant otherwise evaluates your CV against nothing at all.

Here's the trick that does the most work, and it's mildly dishonest: never tell the model the CV is yours. Paste it as "a candidate's CV I'm screening for this role" and the register shifts immediately — the assistant stops protecting your feelings and starts behaving like a reviewer, because nothing in the prompt signals that agreement is wanted. Combine that framing with the three questions above and you'll get the most useful output a chatbot can give you. Then do the unglamorous part yourself: run a structured pass over the document using a 12-point CV self-check so your judgment, not the model's mood, decides what changes.

What can't a chat assistant see in your CV file?

A chat assistant can't see your file — it sees whatever text survived extraction. That single fact invalidates most of the reassurance you get. When you paste your CV, the layout is gone before the model reads a word: two columns collapse into one scrambled stream, text boxes vanish, icons become nothing, graphic skill bars carry no information, and tables flatten into runs of words with no row structure. The assistant then reviews a clean paragraph of text and tells you it's well organized. It is — in the version it received. The version the employer's screening system receives can be a different document entirely, and nothing in the conversation will warn you.

The specific failures are predictable. Contact details placed in a page header or footer are routinely dropped in extraction, so a parser records a candidate with no phone number. Dates parked in a narrow sidebar interleave with job titles from the other column. Creative headings like "What I Bring to the Table" sit where a parser is hunting for "Experience". Acronym-only skills miss keyword searches that look for the spelled-out term as well, so "SEO" alone can lose you a match that "SEO (search engine optimization)" would win. Worst of all, a scanned or image-based PDF contains no extractable text whatsoever — human-readable, machine-invisible, and a chatbot shown a screenshot will still praise it.

Add the gap in knowledge. A chat assistant doesn't know which screening system the employer runs, how that instance is configured, which filters the recruiter applied, or what the knockout questions were. It can't produce a keyword-match percentage, and if you ask for one it will produce a number anyway — invented, confident, useless. This is why visual polish and parsing quality pull in opposite directions so often: the templates that look sharpest are usually the ones built from the elements that break extraction. If your CV was designed rather than typed, see what happens when a beautiful layout meets the scanner before you send it anywhere.

How do you get honest AI resume feedback instead of praise?

You get honest feedback by replacing conversation with scoring. A number against a fixed rubric can't flatter you — it either clears the bar or it doesn't, and it says the same thing on Tuesday as it said on Monday. That's the whole design argument for a scored CV analysis over a chat thread: the same categories every time, applied to the actual file rather than to pasted text, with the weak areas named rather than cushioned. You lose the warmth of a dialogue and gain something you can act on. A 6/10 on clarity and structure tells you where to spend your evening; "this is a strong CV overall" tells you nothing you didn't hope for.

That's the gap HRLens was built to close. Upload a PDF, Word file or even a photo of your CV — Hebrew documents are read with OCR — and the free analysis returns an overall score out of 100 rated Weak, Decent, Strong or Excellent, plus five separate category scores for experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility, along with your real strengths and a visual analysis of the layout. Paste the job description and you'll also see which skills the ad asks for that your CV doesn't mention. You can run a free scored CV check in a couple of minutes and compare it against whatever the chatbot told you.

The honest recommendation is to use both, in order. Draft and rewrite with a chat assistant, because that's where it earns its keep, then score the finished file before you apply — language first, verdict second, never the reverse. If the free score shows a problem you can't place, the paid Full Analysis at $3.99 adds the fix list, the "why you might be rejected" breakdown, your blind spots, an estimate of interviews per ten applications against the market and a salary estimate. Every analysis also generates ATS-friendly rewritten CVs in six templates that mirror the job's keywords. One number you didn't want to hear beats forty agreeable paragraphs.

Frequently asked questions