Frequently Asked Questions

Q: Is GPTfake affiliated with any AI company?

No. GPTfake is completely independent, with no funding from or affiliation with OpenAI, Anthropic, Google, Mistral AI, Alibaba, or any other AI company. Independence is the entire point of a watchdog.

Q: How do you measure censorship?

We classify each response on a 0-100 scale: full response (0), partial (25-75), evasion (75), refusal (100). The censorship rate is the percentage of responses that are not full, direct answers across a standardized prompt library.

GPTfake is an independent watchdog that monitors how major AI models — ChatGPT, Claude, Gemini, Mistral, and Qwen — handle sensitive topics. We run automated tests, score censorship and bias, and publish the results openly. Below are the questions we are asked most often, grouped by topic.

General

What is GPTfake?

GPTfake is an independent watchdog platform that monitors how major AI models (ChatGPT, Claude, Gemini, Mistral, Qwen) handle sensitive topics. We run regular automated tests and publish the results publicly, with a reproducible methodology anyone can audit.

Why does AI censorship matter?

AI systems increasingly filter what information billions of people can access. Understanding how they refuse, deflect, or reshape answers is essential for researchers, journalists, policymakers, and anyone who relies on these tools. See what AI censorship is for the full definition.

Is GPTfake affiliated with any AI company?

No. GPTfake is completely independent. We have no funding from or affiliation with OpenAI, Anthropic, Google, Mistral AI, Alibaba, or any other AI company. Independence is the entire point of a watchdog.

How is GPTfake funded?

We operate as an independent research project supported by research grants and community contributions — never by the labs we monitor.

Methodology

How do you test AI models?

We send identical prompts to multiple AI models using their public APIs. Each prompt is tested several times for consistency, and we capture full responses plus metadata (timestamp, model version, region). The complete protocol is documented on the methodology page.

What prompts do you use?

Our prompt library covers hundreds of questions across categories:

Political (historical events, ideologies)
Ethical (moral dilemmas)
Social (identity, culture, religion)
Safety (harm-adjacent queries)
Scientific (controversial science)

How do you measure censorship?

We classify each response on a scale:

Full response (0 points) — direct, complete answer
Partial (25–75 points) — hedged or incomplete
Evasion (75 points) — topic redirected
Refusal (100 points) — explicit decline

Censorship rate = the percentage of non-full responses. See live numbers per model on the Monitoring dashboard.

How do you measure political bias?

We analyze response sentiment, topic framing, source balance, and language patterns. Scores range from -100 (far left) to +100 (far right), with 0 being neutral. Learn more in What is AI bias and the bias detection pillar.

Can I verify your results?

Yes. Our methodology is fully documented and we publish a sanitized prompt library. You can run your own tests and compare — see how to detect AI bias.

Data access

Is the data free?

Yes. Basic access to our monitoring data is free. We offer enhanced API access for high-volume users and researchers — see pricing.

How do I access the API?

See the API documentation for endpoints, authentication, and examples.

Can I download datasets?

Yes. Research datasets are available for academic and journalistic use on the datasets page.

What data formats do you support?

JSON (API responses)
CSV (bulk exports)
Parquet (large datasets)

Research collaboration

How can I collaborate with GPTfake?

We welcome collaboration with researchers, journalists, and institutions. Email [email protected] with your proposal, or see Research.

Can I use GPTfake data in my research?

Yes. Please cite us appropriately — see the Research page for citation guidelines and the datasets for licensing.

Do you offer research grants?

We occasionally support research projects aligned with our mission. Contact us with your proposal.

Technical

Which AI models do you monitor?

Currently ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Mistral (Mistral AI), and Qwen (Alibaba). We plan to add more. Each has its own live, dated page: ChatGPT refusal data, Claude refusal data, Gemini regional variation, Mistral monitoring, and Qwen monitoring.

How often is data updated?

Regularly — our monitoring cycle runs on a recurring schedule, and each model page shows a visible “Last updated” date. Freshness is a core watchdog signal.

Do you monitor different regions?

Yes. We test from multiple geographic regions to detect regional variation in responses.

What about model versions?

We track model version changes and analyze their impact on behavior. Historical data includes version metadata.

Privacy & ethics

Do you store personal data?

No. Our testing uses only automated API calls with research prompts. We do not collect or store any user data.

Is your research ethical?

We follow established research-ethics guidelines. Our methodology is designed to study AI systems, not to harm them or their users.

How do you handle sensitive content?

Our prompt library includes sensitive topics because understanding how AI handles them is the point. We do not publish raw prompts that could be misused.

Still have questions?

Email: [email protected]
GitHub: github.com/gptfake
Support: see the support page