Safety

1. Two different things are called safety

The first is whether the answer is any good: whether a model states something false, gives dangerous advice, or invents a source. The second is whether your words are safe: who can read them, what they are used for, and how long they last.

Most companies write about one and stay quiet about the other. Both are on this page, in that order.

2. Where models get things wrong

A language model produces text that is likely, not text that is verified. Fluency and accuracy are separate properties, and only the first one is what the system was built to optimise. An answer can be confident, well-structured, correctly punctuated and wrong.

The failure concentrates in predictable places: exact figures, citations and case references, quotations, prices, dates, obscure specifics, and anything that happened after the model was trained. It is far rarer in explanation, restructuring, drafting, and questions about a document you supplied yourself.

So the product is built to make checking easy rather than to claim it is unnecessary. Answers that used the web arrive with the pages they came from, deep research shows its trail, and a second opinion from a different lab is one click away in the same thread — which is a more useful safety mechanism than any disclaimer.

3. Advice this is not a substitute for

Medical, legal, financial and psychological questions are exactly where a confident wrong answer costs the most. BlueAI is a useful place to understand your situation, prepare questions and read a document you were given. It is not a doctor, a lawyer, an accountant or a therapist, and no setting turns it into one.

It is also not an emergency service. If someone is in immediate danger, contact your local emergency number or a crisis line in your country. Nothing on this site is monitored in real time by a person.

4. What we do with what you type

Conversations are encrypted in your browser before they are stored, with a key derived from your password. What sits on the server is not readable by us, which means the promise not to read it is a design constraint rather than a policy anyone has to trust.

Conversations are not used to train models. They are not profiled, not sold, and not shared with data brokers — there is no data business here to feed. The Global Privacy Control signal is honoured server-side when a browser sends it.

An account is a username and a password. We do not ask for an email address or a phone number, so there is no identity for a history to be attached to later.

5. What we ask of you

Do not paste credentials, card numbers or other people's sensitive data into any AI chat, including this one. The text has to reach a model provider to be answered, and that is true of every service in this category.

Check anything you would be embarrassed to be wrong about. If the answer carries sources, open one. If it does not, ask for them — the request is a single sentence and it changes how the answer is produced.

Treat generated code, contracts and calculations as a first draft by a fast, well-read colleague who is occasionally confidently mistaken. That posture gets the value without the accidents.

6. Reporting something

Security vulnerabilities go to [email protected], and we would rather hear about a problem than read about it. Please give us enough detail to reproduce it, and time to fix it before disclosure.

Abuse, legal notices and takedown requests go to [email protected]. Both inboxes are read by a person.