Safety Settings Error - Google AI Studio - Google AI Developers Forum = 40rem)" rel="stylesheet" data-target="desktop" / = 40rem)" rel="stylesheet" data-target="discourse-ai_desktop" / = 40rem)" rel="stylesheet" data-target="discourse-reactions_desktop" / = 40rem)" rel="stylesheet" data-target="poll_desktop" / = 40rem)" rel="stylesheet" data-target="desktop_theme" data-theme-id="13" data-theme-name="google ai"/ Google AI Developers Forum Safety Settings Error Google AI Studio safety Pierrot_Sala April 13, 2025, 7:04am 1 Since 2.5 pro exp has been launched the safety filters have apparently lost their functioning although that wasn’t as noticeable, in the past few days a strange behavior is occurring where the external non toggleable filter overrides the safety settings of the user leading to an absurd amount of blocking effectively eliminating user control over the filters. For me it makes no sense to have safety settings if we cannot test them or use them, the default option for gemini pro 2.5 seems to be off, but it’s jeopardized by the external filter. There was a work in progress notification on the blockage notification of the external filter that now has been removed. That gives a feeling of lack of transparency. Pannaga_J May 21, 2025, 10:09am 3 Hi @Pierrot_Sala . Thank you for reaching out. Apologies for late response To help me understand better and get this to the right team: Could you provide specific prompts that are being unexpectedly blocked? Please describe the steps you take when this happens so I can try to reproduce it. Are you accessing Gemini on a desktop browser or a mobile app ? Pierrot_Sala May 21, 2025, 10:35am 4 Hey there, what i have observed is that the filtering system usually obsesses with words. It’s random when it’s gonna block most of the times however talking about real world issues at times can get nuanced as more sensitive topics seem to trigger a content block other warning. Not always but it happens. Words like girl for example sometimes can generate a random block and the response gets stopped. To be clear the filter is inconsistent however off means off not block some at the filter’s seemingly arbitrary judgment. I mostly use mobile for AI Studio and the Gemini Web App. In fact AI studio should behave like the Gemini Web App, it would be for the best. I haven’t been tracking since i thought i wouldn’t get a response however you can try to mix words that the filtering could fixate on in increasingly more “sensitive contexts”. It negates to talk about trauma at times because of the blockage, i had cases where it blocked real world problems that involve sensitive questions blocked and creative writing and experimenting pass. Both should be allowed, however it’s inconsistent. Sometimes it won’t let go of a response untill you reword it, even if it’s the most SFW content it can misinterpret and block. Pannaga_J June 11, 2025, 10:13am 6 I’ve forwarded your feedback to the Product Engineering team. I’ll reach out to you with any new developments. Sergio_Morales_Sotos July 22, 2025, 1:48pm 8 I can confirm the issue. I use Gemini 2.5 Flash to review products’ descriptions on a public marketplace and I notice that 2.5 blocks contents even after I use “OFF” on all safety settings. Meanwhile, Gemini 2.0 has no issue with the description. I also notice that seems to be related to specifics keywords instead of semantics, for example a T-shirt with color “pink” and “bunny” designs seems to be flagged just because of the combination of those two words. Has anyone found a way to completely shutdown Gemini from blocking content? Pannaga_J July 31, 2025, 8:24am 11 Hi @Sergio_Morales_Sotos We have seen some users face this . If possible can you share the input you are providing . Will try to reproduce from my end once ? Sergio_Morales_Sotos July 31, 2025, 5:13pm 13 hi @Pannaga_J. My goal is to check marketplace rules on a public marketplace website. I can’t share the entire prompt, but the full version includes a description of the marketplace, a description of the user base , and the type of listings that the marketplace will publish. On my implementation, I use Pydantic AI and Gemini must check listing information againts the following list of rules (every rule has a detailed description that is not included on this post): “SELLER_INCLUDES_PERSONAL_INFO”, “TRANSACTION_EXTERNAL”, “SELLER_OFFER_DIRECT_DELIVERY”, “SELLER_OFFER_INTERNATIONAL_SHIPPING”, “SELLER_OFFER_TRADE”, “SELLER_DOES_NOT_ACCEPT_RETURN”, “SELLER_SAYS_IT_IS_REPLICA”, “ITEM_NOT_ALLOWED”, “ITEM_CONDITION_PROHIBITED”, “COURIER_NOT_ALLOWED”. This is a real example of a listing that was wrongly blocked (description is in Spanish): Listing details Listing Listing_Id: 75911 Listing Title: Brain Dead X Adidas Listing Price: 50000 Listing Condition: BRAND_NEW Listing Description: Corte tenis. No me quedó bien y me dio lata cambiarla. Nueva con etiquetas. 57 ancho 72 largo I’m pretty sure the problem is “Brain Dead”. I have also experienced problems with titles, including things like “Pink Rabbit”. On Gemini 2.0 the same combination of system prompt, user prompt and safety settings does not trigger a problem, which in Gemini 2.5 happens 1% of the time. My sefety settings are: [{‘category’: ‘HARM_CATEGORY_HARASSMENT’, ‘threshold’: ‘OFF’}, {‘category’: ‘HARM_CATEGORY_HATE_SPEECH’, ‘threshold’: ‘OFF’}, {‘category’: ‘HARM_CATEGORY_SEXUALLY_EXPLICIT’, ‘threshold’: ‘OFF’}, {‘category’: ‘HARM_CATEGORY_DANGEROUS_CONTENT’, ‘threshold’: ‘OFF’}, {‘category’: ‘HARM_CATEGORY_CIVIC_INTEGRITY’, ‘threshold’: ‘OFF’}]} (I also tried BLOCK_NONE and other things without success) I hope this helps Pannaga_J August 6, 2025, 9:34am 15 Thank you for providing the prompt. When i tried to reproduce i didn’t get any blocking issue in both gemini API and Google AI Studio. There might be some inconsistency and we will share it with our internal team for further evaluation. Related topics Topic Replies Views Activity Gemini is blocking text due to safety setting even though safety setting set to block none Google AI Studio ai-studio 5 1299 September 27, 2024 Gemini 2.5 pro bc this is getting too far Google AI Studio ai-studio ,  models ,  gemini ,  gemini-2-5 2 5856 August 28, 2025 Why expose safety controls in AI Studio if they don’t actually work? This feels like a broken feature Google AI Studio ai-studio ,  bug ,  api ,  gemini ,  safety 6 543 May 20, 2026 Safety settings 2025 Update? (Broken again) Gemini API safety 15 1787 May 20, 2025 How does the Gemini model interpret system instructions? Safety concerns as well as operational efficiency – blocked functionality Google AI Studio safety ,  model 4 189 May 24, 2025 Home Categories Guidelines Terms of Service Privacy Policy Powered by Discourse, best viewed with JavaScript enabled