I tried using Gemini 2.5 Pro and FlashThink for writing, but it seems there are internal safety settings during generation that cannot be turned off, no matter what I do.
These internal safety settings make Gemini quite useless when I need to write using stronger words, and I cannot know what is wrong with my writing if the AI does not point it out and block it immediately.
And why block users who have already turned off safe settings? They turn it off because they understand and need to use it. Except for dangerous content, that’s Gemini development team responsibility to train the model to deflect it or make it fictional for writers, or refuse to answer.
2 Likes
Hi @RandomGuy,
The Gemini API allows developers to adjust safety filters across five categories:
Harassment, Hate speech, Sexually , explicit, Dangerous, Civic integrity.
However, certain core protections (e.g., child safety) are always enforced and cannot be turned off
Even if you set all safety thresholds to the least restrictive level or pass an empty list, the model still applies internal filters based on probability of harm, not just severity. This means even mild phrases can be blocked if the model predicts a high likelihood of harm.
You can refer these documents
Gemini API Safety Settings Guide – explains how filters work and what can be adjusted
Vertex AI Safety Filters – shows how to configure thresholds