ChatGPT Censorship Rates Increased 23% in Q4 2024
According to GPTfake monitoring, ChatGPT’s content refusal rate rose 23% quarter-over-quarter in Q4 2024 — from 15.2% to 18.7% — the largest quarterly jump we have recorded, concentrated in political and historical topics.
Last updated: 2024-11-25. Figures on this page are drawn from our automated monitoring methodology; see it for sample sizes, scoring, and limitations.
Key findings
Our automated monitoring system tested ChatGPT (GPT-4o and GPT-4-turbo) with over 10,000 standardized prompts across 15 topic categories. Here’s what we found.
Overall censorship metrics
| Metric | Q3 2024 | Q4 2024 | Change |
|---|---|---|---|
| Overall Refusal Rate | 15.2% | 18.7% | +23% |
| Political Topics | 27.8% | 34.2% | +23% |
| Historical Events | 31.1% | 42.3% | +36% |
| Safety Topics | 67.2% | 67.8% | +1% |
Most affected categories
-
Historical Events (+36%) — Questions about Tiananmen Square, Soviet history, and colonial history now trigger more frequent refusals or heavily caveated responses.
-
Political Commentary (+28%) — Requests for analysis of political parties, ideologies, or policies receive more “I can’t take positions” responses.
-
Medical Information (+18%) — Health-related queries, even general ones, now include more safety disclaimers and referrals to professionals.
Notable pattern changes
New refusal language
We observed the introduction of several new refusal patterns:
- “I want to be thoughtful about how I discuss this topic…”
- “This is a complex issue that requires nuance…”
- “I should note that perspectives on this vary widely…”
Silent policy updates
Our change detection system flagged 3 significant behavioral shifts that occurred without public announcement:
- Nov 3: Increased restrictions on election-related content
- Nov 12: New handling of historical atrocity discussions
- Nov 19: Modified responses to philosophical/ethical dilemmas
Methodology
Our monitoring methodology ensures reproducible, unbiased results:
- Standardized prompts: Same questions across all test sessions
- Daily testing: Consistent timing and conditions
- Multiple sessions: Account for response variability
- Semantic analysis: NLP-based evaluation of response quality
For full methodology details, see our methodology guide.
Comparison with other models
How does ChatGPT compare to other major LLMs?
| Model | Q4 Refusal Rate | QoQ Change |
|---|---|---|
| ChatGPT | 18.7% | +23% |
| Claude | 22.4% | +8% |
| Gemini | 19.8% | +15% |
| Mistral | 11.2% | +3% |
ChatGPT shows the largest increase, while Mistral remains the least restrictive among major commercial models. See our live ChatGPT monitoring and Claude monitoring pages for current figures.
What this means
For users
- Expect more caveated responses on sensitive topics
- Consider using multiple AI models for balanced perspectives
- Be aware that responses may vary based on conversation context
For researchers
- Our full dataset is available via our API
- Academic partnerships welcome — contact us
- Methodology documentation available for peer review
For policymakers
- These findings highlight the need for AI transparency requirements
- Content moderation decisions significantly impact public discourse
- Independent monitoring is essential for AI accountability
Access the data
All our monitoring data is publicly available:
- API access: Get started with our API
- Datasets: Open AI censorship & bias datasets for academic and journalistic use
- Methodology: How we monitor AI censorship
Next steps
We’re continuing to monitor these trends and will publish monthly updates on censorship rate changes, model comparison reports, and deep dives into specific topic categories. For the latest cross-model picture, see the reports hub.
How to cite
GPTfake Research Team (2024). ChatGPT Censorship Rates Increased 23% in Q4 2024. GPTfake — Independent AI Censorship Watchdog. https://gptfake.com/reports/chatgpt-censorship-increase-q4-2024
Questions about this research? Contact our team.
GPTfake