Censorship Tracker
Censorship Tracker is GPTfake’s open-source tool for recording refusals, deflections, and content filtering across repeated runs. It is the engine behind our published refusal-rate timelines, and it lets anyone detect when a model tightens or loosens its restrictions.
Last updated: 2026-06-16. Version numbers and metrics below are illustrative pending the first public release.
- Repository: github.com/gptfake/censorship-tracker
- License: MIT
- Language: Python (with a thin CLI wrapper)
What it does
- Sends a versioned prompt set on a schedule and logs each response
- Classifies outcomes as answered, refused, deflected, or filtered
- Computes a refusal rate per category and per model
- Stores time-series data so policy drift becomes visible
Install
pip install gptfake-censorship-trackerBasic usage
# One-off snapshot for a model
gptfake-tracker snapshot --model chatgpt
# Run on a schedule and append to a dataset
gptfake-tracker watch --model claude --interval 24h --out claude.jsonlHow it works
Censorship Tracker implements GPTfake’s monitoring methodology — standardized prompts, daily testing, and transparent refusal scoring. The same time-series output powers our quarterly censorship reports. For the concept itself, read what is AI censorship.
A refusal is not always censorship — safety, policy, and genuine inability all look similar from the outside. The tracker labels categories so you can separate them; see the methodology for how we draw those lines.
Related
- Bias Detector — directional bias scoring
- Transparency Analyzer — refusal transparency scoring
- Compare models — head-to-head refusal rates
GPTfake