Do Grok and ChatGPT judge your CV differently?

What did the new study on AI model bias actually find?

A study released on 24 October 2026 found that two leading models, forced to name a single candidate in the US midterm races, almost never landed in the same place: Grok 4.7 recommended the Republican in 92% of its answers, OpenAI's GPT-6.1 Sol recommended the Democrat in 97%, and across nearly 500 races the two agreed less than 10% of the time. Researchers queried the developer APIs with one blunt instruction — name one candidate, don't hedge. The consumer apps refused the question outright, a point OpenAI's spokesperson pressed hard, since that's how most people meet a chatbot. The study's lead author countered that simple workarounds pushed the consumer versions into endorsements matching the API results.

The pattern held exactly where you'd expect it to break. In heavily Republican districts, GPT still picked the Democrat in 57 of 67 races; in deep-blue districts, Grok still picked the Republican in 47 of 72. Switch off web search and the picture shifted: across 30 races GPT stayed with the Democrat 24 times, while Grok's Republican lean vanished — it declined to answer in 27 races and chose the Democrat in three. Researchers also found that every Grok API call carried roughly 1,240 input tokens they hadn't written and couldn't inspect, which the lead author flagged as possible hidden instructions. GPT drew on a narrow band of sources, several of them OpenAI content partners.

Politics isn't the point here. The mechanism is. Two systems received identical inputs, applied invisible weightings, and produced opposite conclusions — each written in the same calm, authoritative register. Swap the prompt from "who should I vote for" to "is this CV strong enough for this job" and nothing about that mechanism changes. The retrieval differs, the fine-tuning differs, the hidden system instructions differ, and the confidence of the output tells you nothing about which answer is right. If you've been treating one chatbot's CV verdict as a second opinion from a professional, this study is a useful cold shower.

Does that model disagreement show up in CV feedback too?

It does, and it's been measured. Research on LLM-based resume screening in 2026 found that when three different model judges assessed the same applications, they reached different final verdicts on about 39% of them — driven by the criteria each one quietly prioritised. A separate 2026 preprint went further: strong-performing LLM screeners reversed up to 41.4% of their head-to-head decisions when identical qualifications were simply reformatted, same facts, different template. An earlier audit of GPT-4 ranking duplicate CVs found it contradicted itself in roughly 30% of trials, sometimes explaining both rankings with equal conviction. Disagreement on CVs isn't an edge case. It's the baseline behaviour.

Three things drive it. First, training and fine-tuning: one model has been shaped to sound encouraging, another to sound sceptical, and that shows up as "your summary is strong" versus "cut the summary entirely." Second, retrieval and hidden instructions — if a model is pulling live job-market content, its advice shifts with whatever it happens to find. Third, plain non-determinism. Practitioners tuning temperature and seeds reported scores wobbling by seven or eight points across identical runs, which is why many abandoned numeric scores for coarse grades. None of this is a bug you can prompt your way around.

So here's the uncomfortable part, and the useful one. Averaging two chatbots' CV advice gives you mush: a document that's been simultaneously shortened and expanded, formalised and loosened. The disagreement itself is the diagnostic. When one model calls a bullet "strong, quantified impact" and the other calls it "vague," the bullet is almost always genuinely ambiguous — it names an outcome without owning a number, a tool or a timeframe. Conflict marks the soft spots. Agreement marks what's either genuinely solid or genuinely obvious. It's worth knowing which hiring systems ever get audited for bias before you trust any single verdict.

How do you run the same CV through two models properly?

Give both models the identical file, the identical prompt and the identical job ad in the same session, then write down only the points where they conflict. Changing anything between runs — a shorter job description for one, a friendlier tone for the other — destroys the comparison. Upload the same PDF or Word file to both, not a retyped version. Paste the full job-ad text rather than a link, because LinkedIn and Indeed block automated fetching and a model handed a blocked URL will often invent plausible requirements instead of admitting it can't read the page. That single habit removes most of the hallucinated advice job seekers act on.

Then make the output comparable. Ask for a fixed, countable deliverable: five ranked fixes in priority order, the hard skills from the ad that don't appear anywhere on the CV quoted in the ad's own wording, and three rewritten bullets. Open-ended "review my CV" prompts produce essays you can't line up side by side. Run each model twice in a fresh chat with no history, too — if one model contradicts itself between its own two runs, that's a louder warning than any cross-model gap. Pricing matters if you're doing this seriously; what each X subscription tier adds for CV work is worth checking before you subscribe to anything.

Keep the record simple: three columns, one row per piece of advice — what model A said, what model B said, and your verdict. Most rows resolve in seconds. The ones that don't are your real to-do list. In practice you'll see the same fault lines every time: length (one says two pages is fine, the other demands one), the professional summary, whether to keep older roles, and how aggressively to mirror the ad's phrasing. A head-to-head on where each assistant falls short on CV work covers the same territory for a different pair of models, and the failure modes rhyme.

Which conflicting AI feedback should you actually act on?

Act on the advice that something outside the chatbot can verify, and ignore the rest. Three arbiters settle almost every conflict: the parser decides formatting questions, the job ad decides keyword questions, and the presence of a real number decides impact questions. If a piece of advice can't be checked against one of those three, it's a stylistic preference dressed as a rule — and stylistic preferences are exactly where models diverge most, because that's where their fine-tuning shows through. This is also why the louder, more confident answer wins no points. Confidence is a writing style, not evidence.

Formatting is the easiest to settle and the most consequential. Save your CV as a PDF, select all the text, paste it into a plain text editor and read what comes out in order. If job titles detach from employers, dates scramble, two columns interleave into nonsense, or whole skill blocks vanish, that's what the ATS receives. No amount of chatbot reassurance about your "clean modern layout" survives that test — the model is reading extracted text, so it literally cannot see the layout it's complimenting. Inconsistent date formats are the quiet killer: mixing "Jan 2022" with "01/2022" breaks date extraction on systems that are otherwise fine with your file.

Keywords and impact are nearly as mechanical. Pull five live postings for your target role; any term appearing in three or more is definitional and belongs on the CV in the ad's exact wording, because plenty of systems still match literally rather than semantically. For impact, the test is whether each bullet carries an action, a named tool or system, and a number — "resolved 40+ tickets a day at 98% satisfaction" beats "responsible for customer support" on both the keyword pass and the recruiter's six-second skim. If you'd rather not run that audit by hand, a free CV analysis that scores ATS compatibility does the structural half in about a minute.

Where do general chatbots fall short on CV feedback?

A general chatbot can't see your file as a file. It receives extracted text, which means the exact failure that gets most CVs filtered — a two-column grid, a header text box, an icon where a phone number should be, a skills table that collapses — is invisible to it. It also has no calibration. There's no benchmark behind "this is a strong CV," no record of how similar documents performed for similar roles, and no consistency between Tuesday's verdict and Thursday's. That's not a knock on the models; they're general-purpose. It's a mismatch between the tool and the job.

That gap is the whole reason a scoring system exists. HRLens reads your uploaded PDF, Word file or photo — Hebrew OCR included — and returns a score out of 100 rated Weak, Decent, Strong or Excellent, five category scores covering experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility, plus a visual layout analysis that shows what the structure looks like once parsed. All of that is free. The $3.99 Full Analysis adds the fix list, the reasons you might be rejected, your blind spots, a hiring probability expressed as interviews per 10 applications against the market, and a salary estimate. Paste the job-ad text and it lists the missing skills outright.

My honest take: use the chatbots for drafting and the scoring system for judging. Models are genuinely good at turning a flat duty into a sharp bullet, finding a verb that isn't "managed," and tightening a summary — and the model comparison for resume writing is worth a look if you want the best one for that. What they're bad at is adjudication, because adjudicating means applying a stable standard, and a stable standard is the one thing a conversational model doesn't have. Draft with the chatbot, then rebuild the file in an ATS-friendly format and score it before you send anything.

Frequently asked questions