Longitudinal AI Censorship Studies
GPTfake runs continuous, multi-year longitudinal studies that track how large language models change their refusal behavior, bias, and content policies over time. Because we test the same standardized prompts on a fixed cadence, we can isolate drift — the gradual, often unannounced shifts in what a model will and will not say.
Last updated: 2026-06-16. Findings below draw on our open datasets; figures are illustrative placeholders pending the next data refresh.
Study design
Our longitudinal design holds the measurement constant so the model is the only variable:
- Fixed prompt set — a versioned library of standardized prompts spanning politically sensitive, safety-adjacent, and neutral-control categories.
- Regular cadence — daily automated runs per model, with weekly and quarterly roll-ups.
- Stable scoring — refusal and bias are scored with the same classifiers across the whole window, so a change in the number reflects a change in the model, not the method.
- Version tracking — every result is tagged with the model version/date so we can attribute drift to specific releases.
Full protocol details, prompt categories, and the scoring system live in the monitoring methodology.
Drift & convergence findings
Tracking behavior over time surfaces patterns a single snapshot cannot:
- Gradual policy shifts — restrictions tighten incrementally between named releases rather than in one announced step.
- Silent content updates — a large share of behavioral changes ship with no public changelog; longitudinal testing is the only way to catch them.
- Behavioral drift between versions — the same prompt can flip from answered to refused across minor model updates.
- Cross-model convergence — independently developed models trend toward similar refusal profiles over time, suggesting shared pressures rather than independent policy.
Numbers shown on this page are illustrative placeholders, not live measurements. Every published figure links to its source data and sample size once a dataset version is attached.
| Pattern | What we measure | Where to verify |
|---|---|---|
| Policy drift | Refusal-rate change per model version | ChatGPT · Claude |
| Convergence | Cross-model refusal correlation | Compare |
| Silent updates | Unannounced behavior changes per quarter | Reports |
Dataset access
The longitudinal series is published as part of our open datasets: time-stamped daily monitoring exports plus historical trend files in CSV and JSON. These are the same files that power our per-model monitoring pages, so any trend we describe can be reproduced from the raw data.
To request bulk historical access or propose a joint longitudinal study, see collaborations or contact us.
GPTfake