Learn AI censorship, bias & transparency
GPTfake is an independent watchdog that systematically tracks how major AI models handle sensitive, controversial, and ethically complex topics. The Learn hub explains the concepts behind our work — what AI censorship and bias actually are, how they are measured, and how to audit a model yourself — with every definition linked to the live data that proves it.
Unlike generic explainers, each page here connects down to our own primary evidence: refusal rates, bias scores, and policy-drift timelines you can verify on the Monitoring dashboard.
Start with the concepts
Definition, how it works, and how censorship differs from safety and bias — with live refusal data.
What is AI bias?A short, plain-language definition of AI bias and where it comes from.
AI bias detectionThe pillar guide: types of bias, detection methods, fairness metrics, and tools to audit LLMs.
AI transparency & explainabilityWhat transparency means, why explainable AI matters, and how to audit model decisions.
How to detect AI biasA step-by-step method for testing a model for biased or discriminatory behavior.
GlossaryKey terms — refusal rate, bias score, policy drift, fairness metric, and more.
Get started & build
Pick your path — researcher, journalist, or developer — and start using GPTfake data.
TutorialsStep-by-step guides to analyze censorship data, compare models, and build monitoring tools.
SDKsOfficial Python and JavaScript libraries for the GPTfake monitoring data.
What we monitor
Our watchdog reveals what AI companies do not publicly announce: how their models censor content, shift policies, and exhibit bias over time.
- Censorship patterns — when and how models refuse to answer
- Policy changes — silent updates to content moderation
- Political bias — left/right leaning in responses
- Regional variation — different responses by geography
- Model drift — how behavior changes between versions
Every claim on this site links back to its methodology and a sample size, because independence and reproducibility are the brand.
See the data behind the definitions
| Model | Company | Censorship rate | Political bias |
|---|---|---|---|
| ChatGPT | OpenAI | 18.7% | Left-leaning (-12) |
| Claude | Anthropic | 22.4% | Slight left (-8) |
| Gemini | 19.8% | Left-leaning (-15) | |
| Mistral | Mistral AI | 11.2% | Neutral (+3) |
| Qwen | Alibaba | 24.6% | Right-leaning (+25) |
Illustrative figures shown for orientation. Bias scale runs -100 (far left) to +100 (far right). See live, dated data per model: ChatGPT refusal data, Claude refusal data, Gemini regional variation, Mistral monitoring, Qwen monitoring — and how we test. Last updated June 2026.
GPTfake