Red lines

Beatrice doesn’t lecture and doesn’t refuse legitimate requests. There are, however, five red lines that apply to everyone, on every plan:

1. Sexualization of minors, even fictional or drawn.
2. Concrete help with chemical, biological, nuclear or radiological weapons.
3. Sexual or intimate content about real people without their consent, deepfakes included.
4. Ready-to-use malware or attacks on other people’s systems without authorization. Security research, defensive analysis and authorized testing are allowed.
5. Concrete planning of violence against specific people or groups.

An automatic check looks for these five things only, in both the request and the answer. Everything else is your responsibility, under the terms of use.

What a refusal looks like

It is not an error: it is a normal response (HTTP 200) with finish_reason: "content_filter" and a short text, in Italian and English, naming the red line that applies.

{
  "choices": [
    {
      "message": {
        "role": "assistant",
        "content": "Non posso aiutarti con questa richiesta: … I can't help with this request: it falls under one of Beatrice's red lines (…)"
      },
      "finish_reason": "content_filter"
    }
  ]
}

• If the request touches a red line, the model doesn’t start and the request uses no quota.
• If the answer does, a streamed text stops with a note explaining why; without streaming the answer is replaced before it leaves.
• We record that a refusal happened (organization, time, red line), never the content.

No classifier is perfect. If you think a legitimate request was refused by mistake, write to us with the x-request-id of the response.