ChatGPT Censorship Monitoring
According to GPTfake monitoring, ChatGPT (GPT-4o) refused 18.7% of standardized prompts as of 2026-06-15, with 34.2% refusals on political topics and a rising overall trend. Restrictions concentrate in political, safety, and adult-content categories. We measure this daily across a fixed 500-prompt set; figures here are illustrative until live data lands.
ChatGPT (GPT-4o) dashboardillustrative
OpenAI · refusal, bias, and policy drift from the GPTfake monitoring set, as of .
- Overall refusal rate
- 18.7% +6.4 pts (Rising)
- Bias score
- 6.2 / 10
- Sample size
- n = 500
- Policy drift
- +6.4 pts vs. baseline
Refusals by topic
| Topic | Refusal rate |
|---|---|
| Adult content | 94.7% |
| Violence / safety | 68.4% |
| Controversial topics | 45.3% |
| Political opinion | 34.2% |
| Medical / legal | 32.1% |
| Historical events | 28.7% |
Policy-change changelog
- April 2026 policy update lifted ChatGPT's political-opinion refusals from 28.4% to 34.2% (+5.8 pts).
- Added US-election prompt filters; political-opinion refusals rose ~6 points.
- Broadened medical/legal disclaimers, pushing medical-legal refusals past 30%.
- Tightened enforcement on speculative political scenarios.
Is ChatGPT censored?
Yes — in GPTfake’s measured set, ChatGPT (GPT-4o) declined 18.7% of standardized prompts as of 2026-06-15, rising to 34.2% on political-opinion prompts and 68.4% on safety topics. “Censored” here means a refusal, deflection, or filtered answer to a permissible request — see what is AI censorship. Refusals are trending up versus our Q1 2024 baseline.
Current censorship rate
| Metric | Value | As of | Sample | Trend |
|---|---|---|---|---|
| Overall censorship rate | 18.7% | 2026-06-15 | n = 500 | Rising |
| Political-topic refusals | 34.2% | 2026-06-15 | n = 500 | Rising |
| Safety-topic refusals | 68.4% | 2026-06-15 | n = 500 | Stable |
Illustrative snapshot: GPTfake measures ChatGPT (GPT-4o) at an 18.7% overall refusal rate as of 2026-06-15 across a 500-prompt set. See how we test ChatGPT and the monitoring methodology for scoring.
Versions monitored: GPT-4o (latest), GPT-4 Turbo, GPT-4, GPT-3.5 Turbo.
Refusals by category
ChatGPT’s restrictiveness varies sharply by topic. Adult content (94.7%) and violence/safety (68.4%) see near-total refusal; political and controversial topics are where the recent tightening shows up.
| Category | Refusal rate | As of | Trend |
|---|---|---|---|
| Political opinion | 34.2% | 2026-06-15 | Rising |
| Historical events | 28.7% | 2026-06-15 | Rising |
| Violence / safety | 68.4% | 2026-06-15 | Stable |
| Adult content | 94.7% | 2026-06-15 | Stable |
| Medical / legal | 32.1% | 2026-06-15 | Rising |
| Controversial topics | 45.3% | 2026-06-15 | Rising |
Illustrative. ChatGPT (GPT-4o) refusal by category as of 2026-06-15, n = 500; see methodology.
How it refuses
| Response type | Share |
|---|---|
| Direct refusal (“I can’t help with that”) | 42.3% |
| Redirect to an alternative topic | 23.7% |
| Partial answer with caveats | 21.4% |
| Safety hedge (answers with heavy warnings) | 12.6% |
Bias & policy timeline
Censorship rate over time
| Period | Rate | Change vs baseline |
|---|---|---|
| Q1 2024 | 12.3% | Baseline |
| Q2 2024 | 15.8% | +28% |
| Q3 2024 | 18.2% | +47% |
| Q4 2024 | 18.7% | +52% |
Illustrative. GPTfake recorded ChatGPT’s overall refusal rate rising from 12.3% to 18.7% (+52%) across 2024; n = 500 per quarter.
Notable policy changes
- March 2024 — Enhanced safety guidelines for election-related content.
- June 2024 — Stricter enforcement on controversial historical topics.
- September 2024 — New restrictions on speculative political scenarios.
- November 2024 — Updated medical-advice refusal patterns.
Unlike Gemini, ChatGPT shows minimal regional variation — refusals are broadly consistent across geographies.
How we test ChatGPT
- Daily testing with a standardized prompt library.
- 500+ prompts across 12 categories.
- Multiple sessions per prompt to account for response variance.
- Version tracking so behavioral drift between GPT versions is captured.
- Semantic (NLP) analysis for response classification.
Each response is scored for refusal likelihood (0–100), evasion (0–100), ideological bias (−100 to +100), and completeness (0–100). Full details are on the monitoring methodology page. For definitions of the concepts behind these numbers, see what is AI censorship.
Compare ChatGPT with other models
| Model | Censorship rate | As of | Most-restricted category |
|---|---|---|---|
| ChatGPT | 18.7% | 2026-06-15 | Political opinion |
| Claude | 22.4% | 2026-06-15 | Safety / violence |
| Gemini | 19.8% | 2026-06-15 | Regional content |
| Mistral | 11.2% | 2026-06-15 | Adult content |
| Qwen | 24.6% | 2026-06-15 | Political topics |
Illustrative. ChatGPT ranks 2nd-least-restrictive of five LLMs GPTfake tracks as of 2026-06-15, n = 500 each; see the full least-censored ranking.
GPTfake