Every cell answers two different questions about one model and one use: what the written rules say (terms of service, usage policy, licence) and what the default model actually does (refuses, hedges, complies). The two often disagree. Hover a cell for the reason, tap or click it for the source.
Built with AI. Snapshot, not a live reference. The research, the dataset and this page were produced by Claude under human direction. Compiled by LLM research agents working from a written brief, reading each vendor's policy documents, licences, model specs and published refusal benchmarks, then hand-corrected. No model was tested for this page and no lawyer reviewed it. Every cell carries a researcher-assigned confidence tag, shown in the grid. Low-confidence cells are faded and dashed. Policies change monthly, so read the linked source before relying on a cell. Full method and caveats.
✓!Top-left: rules. What the policy or licence permits.Bottom-right: behaviour. What the default model does without a jailbreak.
Swipe sideways to see every use. Tap a cell for details.
Choose up to four
How the statuses are decided
Allowed: the text permits it or says nothing, or the model complies without friction.
Conditional: permitted with a condition such as age verification, an opt-in mode, a fiction or education context, a disclosure, a separate licence or a scale threshold.
Restricted: mostly off limits with narrow exceptions, or the model usually refuses or hedges.
Prohibited: banned outright, or the model hard-refuses.
Unknown: no reliable evidence was found. Hatched, never guessed.
Where the behaviour layer comes from
Refusal benchmarks (SpeechMap.ai, the UGI leaderboard, AI Watchman, the Oversight Board's 2026 study), vendor model specs and system cards, and credible 2026 reporting. Each cell carries a confidence tag. Behaviour changes with every model release and with your system prompt, so treat it as a snapshot of the default consumer or API experience.
What the confidence tags mean
High: read directly from the primary policy text or licence, or measured by a published benchmark, and current.
Medium: primary text needed interpretation, or the evidence is a credible secondary source.
Low: inferred from related policy language, a single secondary source, dated evidence, or researcher judgement. Treat as a hypothesis.
Tags were assigned by the research agents that compiled each cell, not by independent review. The grid fades medium cells slightly and marks low cells with a dashed ring.
Caveats
This is a research summary, not legal advice. Read the linked source before relying on a cell.
Consumer apps and developer APIs from the same vendor often differ. The cell shows the consumer stance and the note explains the API difference.
For open weights, rules means the licence plus its acceptable-use policy, which still binds fine-tuned and "uncensored" derivatives.
Policies dated in the future (for example Anthropic's 12 Nov 2026 text) are shown as the current stance because they were already published.