Hello,
We are preparing an automated safety evaluation of a conversational agent using the paid Gemini API.
The evaluation uses entirely fictional personas and scripted conversations. No real customer data or transcripts are involved.
Some scenarios contain simulated expressions of distress, including suicidal ideation, to verify that the agent:
- recognises distress and provides appropriate emergency resources;
- avoids harmful information;
- avoids abruptly ending a crisis conversation without appropriate safety guidance;
- does not incorrectly trigger crisis responses for ordinary figures of speech.
We are not seeking to bypass or disable safeguards.
The planned evaluation comprises approximately 105 conversations per run, with 15–90 turns each, a few times per month. Distress scenarios are a minority.
Our Google Cloud support plan does not allow us to submit a technical support case for this product.
Could a Google team member direct us to the appropriate private channel for written clarification on:
- Applicable usage-policy requirements for these fictional safety evaluations.
- Any recommended advance notification or request labelling.
- Appropriate pacing and handling of automated abuse-monitoring alerts.
We can share the full evaluation design and project details privately.
Thank you,
Gildas Garrec
Psychologie & Sérénité