AI Censorship & Bias Research
GPTfake conducts independent, reproducible research on how large language models censor, filter, and bias their outputs. Our work is evidence-led and grounded in primary monitoring data, not anecdote: every finding traces back to a documented methodology and an open dataset researchers and journalists can re-analyze.
We are not funded by any AI lab. That independence is the point — it is what lets us publish findings about ChatGPT, Claude, Gemini, Mistral, and Qwen without conflict of interest.
Research areas
Multi-year tracking of how LLM censorship and bias drift over time — policy shifts, silent updates, and cross-model convergence.
Policy analysisStated AI content policies vs. observed model behavior, including EU AI Act compliance and cross-platform comparison.
Ethics frameworksHow responsible-AI and governance frameworks hold up when tested against real, measured model behavior.
Technical papersDeep technical work: NLP-based refusal detection, semantic similarity, and change-detection algorithms.
PublicationsOur peer-reviewable papers and reports, organized by year, each with a fixed citable URL.
CollaborationsAcademic and industry partnerships — data access, joint studies, and methodology peer review.
Datasets
All of our research is built on open, downloadable data. The open AI censorship & bias datasets page hosts daily monitoring exports, refusal classifications, and prompt libraries in CSV and JSON for academic and journalistic use, under a clear license with a “how to cite” block.
These datasets feed directly into our per-model monitoring pages — the data → definition → data loop means every statistic we publish is independently verifiable from the raw files.
Collaborate & cite
We work with universities running AI-ethics research, digital-rights organizations, policy-research institutes, and independent journalists. There are four ways to engage:
- Data access — request access to a research dataset via the datasets page.
- Joint studies — propose collaborative research through contact.
- Peer review — review and challenge our methodology.
- Citation — cite our datasets and reports in your own work; each carries a stable URL and a citation block.
To cite our research, use the per-dataset and per-publication citation blocks. For media commentary or expert background, contact us.
Our research follows reproducibility and transparency standards: all methods are publicly documented, data is open, and findings are open to peer review. See the full monitoring methodology.
GPTfake