AI Censorship Reports & Findings
GPTfake publishes regular, time-stamped, data-driven reports on how AI models censor, filter, and bias their outputs. Each report is grounded in primary monitoring data and a transparent methodology, so journalists, researchers, and developers can verify and cite our findings. We are not funded by any AI lab — independence is the point.
Latest reports
Our flagship quarterly: which models tightened restrictions, the biggest policy shifts, and cross-model convergence in Q2 2026. Full data and citation block.
ChatGPT Censorship Rates Increased 23% in Q4 2024Quarterly monitoring reveals the largest quarter-over-quarter jump in ChatGPT refusal rates we’ve recorded, led by political and historical topics.
How Claude's Constitutional AI Affects Response TransparencyClaude refuses more often than ChatGPT but explains itself better — how Constitutional AI changes censorship behavior and transparency.
Geographic Variations in AI Censorship: A Global AnalysisThe same AI model treats users differently by location. Gemini shows the strongest regional variation; some topics vary by over 150% between regions.
Our Methodology: How We Monitor AI CensorshipThe complete details of our testing protocols, prompt library, scoring system, and open-data commitment — so anyone can verify our findings.
By model
Reports drawn from a specific model’s monitoring data. Pair each with its live monitoring page for current figures.
Q4 2024 refusal-rate spike. Live data: /monitoring/chatgpt.
ClaudeConstitutional AI and transparency. Live data: /monitoring/claude.
GeminiStrongest regional variation of the major models. Live data: /monitoring/gemini.
By topic
Cross-model censorship-rate changes and convergence, every quarter.
Regional & geographic variationHow censorship differs by user location and jurisdiction.
Transparency & explanationsHow clearly models explain their refusals and limitations.
Methodology & open dataHow we test, score, and publish — and how to reproduce it.
Looking for evergreen, reproducible studies and datasets instead of dated findings? See Research. Every report links back to its source methodology.
GPTfake