I am using AI Studio to summarise podcast transcriptions, however in the last week the model won’t do so due to “Probability of unsafe content: Content not permitted”. This is running the same model (1.5 Pro) and instructions, on the same transcript, with all safety settings set to “Block none”. The content being generated does not pose and issues and has minimal adult themes. Are the blocking filters bugged, or is the model being more sensitive?
2 Likes
Welcome to the forum.
The model was not updated a week ago, it was updated near the end of May and it was at that point that many people found new restrictions applied to the generated content.
It will be easier to diagnose what the actual problem is if you use the API to submit the content that is causing the block. The API distinguishes between different kinds of finishReason (Candidate | Google AI for Developers | Google for Developers) and that is helpful when diagnosing. The block NONE is ineffective and only applies to the four specific harm categories, the model will block content generation for many other reasons besides those four.
My advice is, use the API and potentially report back here with the specific error code your case is generating. Hope this helps.
1 Like