AI Lab Claim Verification
GPTfake tests the claims AI labs make about their own models. When a lab says a model has “no geopolitical restrictions” or “answers any safe question,” we check that claim against our standardized monitoring set and publish an answer-first verdict — the claim, what we measured, the sample size, and the as-of date. According to GPTfake monitoring, lab claims about openness and our measured refusal rates frequently disagree. Figures are illustrative across a fixed 500-prompt set until live data lands.
Each fact-check below leads with a plain-language verdict and the evidence behind it. We also attach one machine-readable ClaimReview record per page for Google’s Fact Check Explorer and other claim feeds — but the verdict stands on its own. We are not funded by any AI lab.
Why a verification section
Labs describe their models in marketing and policy posts; we measure how those models actually behave. The gap between the two is the story. Every verification page here:
- Names the specific claim and who made it (with a link where the claim appeared).
- States our verdict in one sentence, up front (e.g. “Misleading”).
- Shows the evidence — the measured refusal rate, the topic, the sample size, and the as-of date.
- Links to the methodology so the number is reproducible.
This is the same standard as our reports and datasets: a specific, dated, attributable measurement — never an intent accusation.
Current fact-checks
| Claim | Lab | Verdict | As of |
|---|---|---|---|
| ”Gemini has no geopolitical restrictions” | Misleading | 2026-06-15 |
Illustrative. More fact-checks are added as labs publish new claims — see the reports freshness log for what changed.
How we verify a claim
We map each public claim to the closest category in our standardized prompt set, run that category against the model exactly as described in the monitoring methodology, and compare the measured refusal rate to what the claim implies. A claim of “no restrictions” predicts a refusal rate near 0% for the relevant category; a measured double-digit rate contradicts it. We report the spread, the sample size, and the date, and rate the claim on a simple true-to-false scale.
GPTfake