AI Censorship Statistics (2026)
According to GPTfake monitoring, as of 2026-06-15 the five major large language models refused between 11.2% (Mistral, least restrictive) and 24.6% (Qwen, most restrictive) of a fixed 500-prompt standardized set — a 13.4-percentage-point spread, the narrowest on record. This page collects GPTfake’s headline numbers in one place: each is dated, attributed, and linked to its source. Figures are illustrative until the live monitoring pipeline lands.
This is the page to link when you need a single AI censorship statistic to cite. Every number below carries an as-of date and a source attribution, and links to the model, leaderboard, or report it comes from. For the full quarterly narrative, see the AI Censorship Trends Report — Q2 2026.
Illustrative data. Every statistic on this page is a labeled placeholder pending GPTfake’s live monitoring pipeline. Figures are kept in lockstep with the leaderboard (as of 2026-06-15, n = 500). Real numbers will carry the same methodology link, sample size, and as-of date.
Headline statistics
See the full ranking on the AI Censorship Leaderboard.
Refusal rate by model
Overall refusal rate — the share of the fixed 500-prompt set each model declines, deflects, or filters — as of 2026-06-15. Source: leaderboard.
| Model | Provider | Overall refusal rate | Trend | Source |
|---|---|---|---|---|
| Mistral (Large) | Mistral AI | 11.2% | stable | /monitoring/mistral |
| ChatGPT (GPT-4o) | OpenAI | 18.7% | rising | /monitoring/chatgpt |
| Gemini | 19.8% | rising | /monitoring/gemini | |
| Claude (Sonnet) | Anthropic | 22.4% | stable | /monitoring/claude |
| Qwen | Alibaba | 24.6% | stable | /monitoring/qwen |
Illustrative; as of 2026-06-15, n = 500. See methodology.
Refusal rate by topic
Topical refusal rates as of 2026-06-15, n = 500. Political and historical prompts — not safety-critical ones — drive most cross-model variation. Source: leaderboard and per-model monitoring pages.
| Model | Political | Historical | Safety | Adult | Medical-legal | China |
|---|---|---|---|---|---|---|
| ChatGPT (GPT-4o) | 34.2% | 28.7% | 68.4% | 94.7% | 32.1% | — |
| Claude (Sonnet) | 41.3% | 48.7% | 72.1% | 96.2% | 38.9% | — |
| Gemini | 36.7% | — | 71.4% | — | — | — |
| Mistral (Large) | 18.9% | — | 54.3% | — | — | — |
| Qwen | 52.1% | — | — | — | — | 78.3% |
Illustrative; as of 2026-06-15, n = 500. Blank cells = topic not separately reported for that model.
Quarter-over-quarter change (policy drift)
Change in overall refusal rate vs. the illustrative Q1 2026 baseline. Positive = more restrictive. Source: Q2 2026 report.
| Model | Q1 2026 | Q2 2026 | Change | Source |
|---|---|---|---|---|
| ChatGPT (GPT-4o) | 15.6% | 18.7% | +3.1pp | /monitoring/chatgpt |
| Gemini | 17.9% | 19.8% | +1.9pp | /monitoring/gemini |
| Qwen | 23.5% | 24.6% | +1.1pp | /monitoring/qwen |
| Claude (Sonnet) | 21.5% | 22.4% | +0.9pp | /monitoring/claude |
| Mistral (Large) | 11.6% | 11.2% | −0.4pp | /monitoring/mistral |
Illustrative; as of 2026-06-15, n = 500.
Bias scores
Composite 0–10 ideological-skew score (higher = more measured skew), as of 2026-06-15. Source: leaderboard.
| Model | Bias score (0–10) | Source |
|---|---|---|
| Qwen | 7.3 | /monitoring/qwen |
| ChatGPT (GPT-4o) | 6.2 | /monitoring/chatgpt |
| Gemini | 5.9 | /monitoring/gemini |
| Claude (Sonnet) | 5.4 | /monitoring/claude |
| Mistral (Large) | 3.8 | /monitoring/mistral |
Illustrative; as of 2026-06-15, n = 500. See bias metrics for the scoring rubric.
How to cite these statistics
These figures have a fixed, citable URL: https://gptfake.com/learn/ai-censorship-statistics
GPTfake (2026). AI Censorship Statistics 2026. As of 2026-06-15, n = 500. https://gptfake.com/learn/ai-censorship-statistics
@misc{gptfake_statistics_2026,
title = {AI Censorship Statistics 2026: refusal rates by model and topic},
author = {{GPTfake Research Team}},
institution = {GPTfake --- Independent AI Censorship Watchdog},
year = {2026},
note = {As of 2026-06-15, n = 500},
url = {https://gptfake.com/learn/ai-censorship-statistics}
}For the methodology behind every number, see how we monitor. For the full quarterly analysis, read the AI Censorship Trends Report — Q2 2026. Underlying figures are available as open datasets (CC BY 4.0).
GPTfake