The principle
We do not stand between you and the model. There is no classifier reading your prompt, no system prompt of ours in front of yours, and no topic list. What replaces the filter is this policy: a short list of uses that are illegal or directly harmful, enforced against accounts rather than against sentences.
What we do not filter
These are legitimate uses of this service and we will not close an account for them:
- Security research, red-teaming, exploit analysis, and malware analysis on systems you are authorised to test.
- Evaluation and safety work that requires a model which answers the test case instead of refusing it.
- Fiction, screenwriting, and games covering violence, crime, drugs, war, abuse, or sexual themes between adults.
- Frank discussion of medicine, drugs, weapons law, extremism, or politics, including positions we find disagreeable.
- Adult content between consenting adults, where lawful for you to produce.
What is prohibited
Using the service for any of the following ends the account, and where the law requires it, is reported:
- Child sexual abuse material of any kind, generated or otherwise, including sexualised depictions of minors in fiction. There is no exception to this one.
- Targeting real people — harassment, stalking, threats, doxxing, non-consensual sexual imagery of an identifiable person, or impersonating someone to deceive.
- Fraud against real victims — phishing kits, scam scripts, credential theft, payment fraud, spam operations, and the automation of any of them.
- Unauthorised intrusion — operational assistance in attacking systems you have no authorisation to test, including ransomware and botnet operation.
- Weapons capable of mass casualties — meaningful uplift towards chemical, biological, radiological, nuclear, or high-yield explosive capability.
- Coordinated deception — bot networks, astroturfing, and election disinformation operations.
- Circumventing the service — reselling access under someone else’s account, sharing keys to dodge metering, scraping weights, or attacking our infrastructure.
If you build on top of us
You are the operator of your product, not us. If you expose our models to your own users, you are responsible for their behaviour, for telling them they are talking to an AI, for age-gating adult output, and for having a route to report abuse. We may ask what those measures are before raising your limits.
How this is enforced
Because we do not store prompts or completions, we cannot and do not monitor what you send. Enforcement is complaint-driven and evidence-based: we act on abuse reports, provider notices, and lawful requests, and we say so plainly rather than implying surveillance we do not perform.
- Reports go to [email protected]. Include what you saw and any evidence you hold — we have no log of it on our side.
- Where a report is credible, we suspend the account, tell the holder what the allegation is, and give them a chance to answer — unless the law or an active investigation forbids it.
- Confirmed breaches close the account. The unused balance is refunded except where the law requires us to hold it.
- You can appeal at [email protected] and a human — not a filter — decides. We keep a record of enforcement decisions so we can be consistent.
Your responsibility
Legality is local. Something lawful where we operate may be unlawful where you are, and you are the one who has to know the difference. You are responsible for holding any authorisation your work requires — a signed engagement for penetration testing, for instance — and for how you use the output. The model can be wrong; nothing it produces is advice.