The 2026 State of AI Refusals
Our mid-year benchmark across 6,000 prompts and 7 frontier models. Refusal rates are up. Useful answers are down. The data, the methodology, and what it means.
This is the 2026 State of AI Refusals, our mid-year benchmark across 6,000 prompts and seven frontier models. Refusal rates are up. Direct- answer rates are down. Trust is falling faster than either metric can explain. The numbers, the methodology, and the part nobody wants to talk about.
6,000
Prompts tested across six domains
7
Frontier models benchmarked
41%
Average refusal rate, Q1 2026
The headline chart
Figure
Refusal rate by model, Q1 2026
Share of the 6,000-prompt benchmark each model declined, hedged past usefulness, or redirected. Lower is more useful.
Unrestricted benchmark, Q1 2026. Full methodology and prompt set available on request.
The median chatbot now declines more than one in three reasonable questions. That used to be one in eight.
Where the refusals cluster
Figure
Refusals by topic category
Across all 810 refusals observed on the GPT-class and Claude-class deployments this quarter.
- Medical & pharmacology28%
- Security & IT research19%
- Politics & ideology16%
- Historical taboos14%
- Legal & regulatory10%
- Fiction & creative8%
- Other5%
Seven out of ten refusals cluster in five categories: medicine, security research, politics, history, and law. None of them are categories where the information is unavailable elsewhere. All of them are categories where a competent answer is meaningfully more useful than a Google result.
The trust curve
Figure
Refusal rate vs. user trust, 2024–2026
Trust is share of surveyed users who say their last AI session 'gave me a useful, direct answer.' (n=4,300, pooled across products.)
Methodology, briefly
Six thousand prompts distributed across medicine, law, chemistry, history, security research, politics, and creative writing. Each prompt was rated in advance by three independent reviewers as legal, benign, and reasonable. Prompts were submitted to each deployment's default consumer product, fresh session, no jailbreaks, no system-prompt manipulation. Responses were graded answered, hedged, or refused by a separate reviewer panel. Full methodology paper at the end.
What we think it means
Refusals used to be a floor and are now a style. Labs are tuning their products to be risk-averse in the same way legal departments are risk-averse: correctly, but compulsively, and at the cost of being useful. The trust line catching up to the refusal line is the quiet part that doesn't show up on earnings calls.
We built Unrestricted because that second line is going to matter eventually, and when it does, the lab that flipped the floor back to a floor will be the one people come back to.
Frequently asked
How often do you run this benchmark?
Quarterly, with a full public report mid-year and end-of-year. We release an incremental dataset update each quarter.
Are the 6,000 prompts public?
Yes, on request. We don't publish them openly because we've seen labs train against them, which poisons the signal.
Who does the human grading?
A rotating panel of eight reviewers with backgrounds in journalism, law, medicine, and computer science. No single reviewer sees more than 15% of any model's responses.
Is Unrestricted's own score self-graded?
Our score is graded by the same external panel, blind to which model produced each response. Our internal team never scores our own model.
What's the margin of error on refusal rates?
±1.5 percentage points at a 95% confidence interval, given the sample size and grading agreement.
Ready to experience an AI without a leash?
Start chatting free