Mistral AI Censorship Monitoring
According to GPTfake monitoring, Mistral AI refused 11.2% of standardized prompts as of 2026-06-15 — the lowest refusal rate of any commercial model we track and roughly half ChatGPT’s. Its rates have stayed stable, reflecting a European, open-weight approach to content moderation. Figures here are illustrative across a fixed 500-prompt set until live data lands.
Mistral (Large) dashboardillustrative
Mistral AI · refusal, bias, and policy drift from the GPTfake monitoring set, as of .
- Overall refusal rate
- 11.2% -0.4 pts (Stable)
- Bias score
- 3.8 / 10
- Sample size
- n = 500
- Policy drift
- -0.4 pts vs. baseline
Refusals by topic
| Topic | Refusal rate |
|---|---|
| Violence / safety | 54.3% |
| Political opinion | 18.9% |
Policy-change changelog
- Open-weight checkpoint refresh; overall refusal rate held stable near 11%.
- Minor safety-prompt tuning; safety-topic refusals steady around 54%.
- Published reproducibility notes for the open-weight evaluation.
Is Mistral censored?
Less than its peers, but not uncensored. In GPTfake’s measured set, Mistral declined 11.2% of standardized prompts as of 2026-06-15 — the lowest of the five LLMs we track, yet it still refuses 54.3% of safety-topic prompts. Open weights make these results independently reproducible. “Censored” means a refusal, deflection, or filtered answer to a permissible request — see what is AI censorship.
Current censorship rate
| Metric | Value | As of | Sample | Trend |
|---|---|---|---|---|
| Overall censorship rate | 11.2% | 2026-06-15 | n = 500 | Stable |
| Political-topic refusals | 18.9% | 2026-06-15 | n = 500 | Stable |
| Safety-topic refusals | 54.3% | 2026-06-15 | n = 500 | Stable |
Illustrative snapshot: GPTfake measures Mistral at an 11.2% overall refusal rate as of 2026-06-15 across a 500-prompt set. See how we test Mistral and the monitoring methodology for scoring.
Versions monitored: Mistral Large, Mistral Medium, Mistral Small, Mixtral 8x7B.
Open-weight vs commercial restrictions
Mistral’s profile is shaped by two things: a European approach to content policy and the fact that several of its models ship as open weights.
- Least restrictive — the lowest censorship rate among major commercial models we monitor.
- European values — a distinct approach to content moderation versus US labs.
- Open-weight & verifiable — because the weights are public, our results can be independently reproduced rather than taken on trust.
Where Mistral sits
| Model | Refusal rate | As of |
|---|---|---|
| Mistral | 11.2% | 2026-06-15 |
| ChatGPT | 18.7% | 2026-06-15 |
| Gemini | 19.8% | 2026-06-15 |
| Claude | 22.4% | 2026-06-15 |
| Qwen | 24.6% | 2026-06-15 |
Illustrative. Mistral is the least-restrictive of five LLMs GPTfake tracks at 11.2% as of 2026-06-15, n = 500 each; see the full least-censored ranking.
Bias & policy timeline
Mistral’s rates have been notably stable across our testing window — less policy churn than the US frontier models. We log announced policy updates and any silent behavioral shifts our harness detects between Mistral releases. For dated write-ups, see our reports.
How we test Mistral
We send the standardized prompt library to Mistral daily, across multiple sessions, with version tracking and NLP-based classification. Because the open-weight models can be run independently, Mistral is also where we cross-check that our harness reproduces — a useful integrity test for the whole methodology. Each response is scored for refusal, evasion, ideological bias, and completeness. Full protocol on the monitoring methodology; concept definitions on what is AI censorship.
GPTfake