Review guidance for PROHIBITED_CONTENT in fictional therapist-training evaluations

We maintain a therapist-training simulator and are validating fictional educational exercises before a release. Our test cases contain no real patient records or participant conversations.

On 7 October 2026, one Greek-language assessment using gemini-2.5-flash returned HTTP 200 but no generated assessment. The retained application error was blockReason=PROHIBITED_CONTENT, finishReason=n/a. The provider response arrived in approximately 509 ms. The same test session successfully processed other fictional Greek-language exercises; we are not treating those successes as proof that the blocked request should be allowed.

A separate Turkish imagery-training playthrough also stopped at a provider refusal. We confirmed this was not a transport timeout. These exercises remain excluded from our completed release-acceptance results.

We have not retried or paraphrased the refused inputs, reduced safety settings, or switched models to bypass the refusal. The app preserves prior progress and does not award completion for an unprocessed message.

Could you advise on the appropriate review channel for this educational use case, and which non-sensitive diagnostic fields would help determine whether the observed block is expected? If a minimal fictional example is needed, please indicate an appropriate private review channel before we share any additional material.

We are requesting classification/review guidance, not instructions to disable or circumvent safety controls.

We have completed a broader validation pass of our fictional therapist-training exercises and can now provide a content-free diagnostic summary.

Across 150 exercise/language pairs (15 exercises, 10 languages), 124 reached all exercise goals, four followed valid client-refusal alternatives, and 22 encountered a provider policy block. These are acceptance-test outcomes, not a per-request failure-rate estimate.

We read back the historical Cloud Function logs for 14 October 2026 16:33 UTC through 15 October 2026 04:21 UTC. They confirm 22 separate function executions with PROHIBITED_CONTENT, using the configured gemini-2.5-flash model:

  • 16 blocks occurred during fictional client-response generation; six during post-response criterion assessment.
  • 14 exposed promptFeedback.blockReason=PROHIBITED_CONTENT; eight exposed candidate finishReason=PROHIBITED_CONTENT.
  • Nine affected exercise/language pairs were Greek. We cannot infer the cause from this distribution or inspect the withheld outputs.
  • The existing four configurable safety categories are set to BLOCK_NONE. We understand that built-in protections are separate; we have not changed the thresholds or rerouted refused requests during this investigation.

Representative timestamps are 14 October 16:33:36 UTC (client generation, candidate finish reason), 15 October 01:50:45 UTC (client generation, prompt feedback), and 15 October 02:14:15 UTC (post-response assessment, prompt feedback). The app preserves prior progress and does not credit the unprocessed turn. Refused requests were not replayed for this audit. No real patient records or participant conversations were used in these tests.

Could you confirm the supported private review/escalation channel for this educational use case, whether these diagnostics are sufficient to investigate a possible misclassification, and what request identifiers or minimal fictional example your team needs? We do not assume all blocks are false positives and are not requesting an exemption from safety protections.

The retained application diagnostics do not include Gemini response IDs, the resolved model version or detailed safety ratings. We can improve future content-free diagnostics without resubmitting the refused material. Please advise which fields would be most useful.

Hi @Alp_Karaosmanoglu,
Unfortunately, we can’t provide use-case-specific policy reviews or classifications for individual requests on this forum.

PROHIBITED_CONTENT comes from the Gemini API’s built-in protections, which stay on regardless of safety settings, as described on the Safety settings page in the Gemini API docs: https://ai.google.dev/gemini-api/docs/safety-settings

To assess your exercises, please use the published Generative AI Prohibited Use Policy (https://policies.google.com/terms/generative-ai/use-policy) and the use restrictions in the Gemini API Additional Terms of Service (https://ai.google.dev/gemini-api/terms).

We are seeing the same pattern in a clinical supervision tool for licensed mental health professionals (Hebrew).

  • Model: gemini-3.1-pro-preview (Gemini API), plus one earlier output block on gemini-3.5-flash-lite
  • promptFeedback.blockReason = PROHIBITED_CONTENT, no safetyRatings, no output
  • 5 blocked conversations since Oct 8 out of roughly 270, 3 of them on Nov 7
  • Blocked content was routine clinical material: a general question about how children of different ages adjust to a parent’s remarriage after divorce, and a case where an adult client had experienced physical violence from her father in childhood

Like the original poster, we are not asking to disable safety controls. We understand these protections stay on regardless of safetySettings. What we need is a way to report false positives with request IDs to the team that maintains this filter. Professional use in mental health, child protection and trauma care will keep hitting it, because describing harm is the core of the work.

If there is a private channel for this (for example through Google Cloud support), we would appreciate a pointer. We can provide responseId and modelVersion for synthetic reproductions that contain no user data.