False-positive PROHIBITED_CONTENT block on legitimate legal-professional content

Gemini API
1

We run a legal research SaaS for Turkish lawyers (Celse AI). Our chat feature summarizes and analyzes real criminal case facts on behalf of licensed attorneys preparing a defense — this necessarily includes describing allegations of fraud, violence, and personal/relationship details as recounted in actual court testimony.

A legitimate query describing a client’s personal relationship in the context of a fraud defense (no sexual content generated, no minors involved, purely factual legal analysis) was blocked with blockReason: PROHIBITED_CONTENT on gemini-flash-latest. We understand this reason isn’t adjustable via safetySettings, but the underlying content is standard criminal-defense material any licensed attorney would discuss.

Is there any guidance for legal-professional/document-analysis use cases that legitimately need to process this kind of factual (non-graphic, non-generated) sensitive content? Happy to share a redacted example if useful.

3

Hi @iber_technology,
The finish_reason: SAFETY response occurs when the input prompt or model response crosses the safety threshold for one of the evaluated harm categories (e.g., harassment, hate speech, sexually explicit content, or dangerous content).

When processing court records or criminal case filings, legal terminology can inadvertently trigger default safety filters. For legitimate professional use cases, you can adjust the safety thresholds directly in your API request using safety_settings.

We cannot provide specific guidance on your use case on this forum.