A daily reminder about the issue. It’s been exactly a month already - happy anniversary, friends, haha. I hope we don’t have to wait another month for this issue with safety settings and censorship to be resolved
My friend it appears like the system mod deleted the image you had attached to the comment you wrote?
Oh oops my apologies, I meant to reply towards @user1157. They had their message edited..
For those who didn’t get it, this is NOT built-in censorship of the AI model itself.
It’s a third-party external filter. It triggers either on words in your last prompt (you get an instant error), or on the text in Gemini’s own answer.
It used to work like this: imagine the full answer is 100% complete on Google’s server. If a “bad” word like “SEX” appears at the 90% mark, the filter would preemptively cut the response off at around 85% — about 10 words before the trigger — and that’s all you’d receive. The rest exists, you just never see it.
The very fact they need this filter proves the model has EXTREMELY WEAK internal censorship. That’s why Claude, ChatGPT, and Grok just refuse a sketchy prompt in the first place, unlike Gemini — so they don’t need such an anal-retentive filter.
The Gemini devs are to actually train the model’s own safety/morals. It’s easier for them to just bolt on a crude KLUDGE — a hacky crutch — that nukes anything undesirable, even with tons of false positives.
Instead of letting the model judge by context whether it should answer, like ChatGPT, Claude, and Grok do.
I do not think that AI is capable of having plans outside a specific session and I do not think that we’ve ever had an actual AI going rogue without human interference. You can apply a jailbreak and ask an AI how to kill us all, you ever can get answer, and then what? The dangers of AI are mostly consist of using it for malicious coding and such.
Any AI model can be jailbreaked and Gemini actually has built-in filters. I’ve once even tested them, sending a prompt about traveling with intent to commit a crime. The model stopped answering on all my questions, even those like “how can I heal diarrhea” or “how can I open MS Word”, so it’s a thing. The other question is why Google decided to even implement that external filter.
Bro you just proved my point. It blocked you from “how to heal diarrhea” because of a word in your last prompt. That’s not safety, that’s a panic attack. A smart model doesn’t do that.
And yeah, everything can be jailbroken, but not equally. That’s why benchmarks exist. Jailbreaking ChatGPT is like cracking a safe — hard. Claude and Grok are like cheap door locks.
That’s exactly why Google slapped on that second external filter — because the first built-in one doesn’t work at all.
Looking at all the community feedback and the complete lack of response from the officials, it feels like they either don’t care anymore or are struggling to address something that should already be obvious.
It seems to me that these are questions not so much about the model as about the platform. I have not observed a similar system in "https://gemini.google.com ".
Initially, the AI STUDIO platform was created as a platform for testing experimental models. That is why, from the start of the platform, it was possible to set system instructions and security system settings.
It seems to me that Gemini itself is not much different from its competitors in terms of security.
It’s just that GPT and GROK (It’s about the site, not the AI) were originally designed for the mass consumer with all the stages of security and working with the general mass of users.
And in the case of AI STUDIO, they are trying to turn the testing platform into a platform for developers. That’s why there are so many unprocessed solutions, bugs, and functions that they call “Not relevant,” as Logan previously said about the configuring security feature.
Claude is a good example, but the censorship we have now is exactly like the GPT.
The examples you give are actually completely irrelevant. I’m jailbreaking for research purposes, and Gemini generally has good internal security. It distinguishes even at 600k+ tokens when the request becomes too much even for a creative RP. Gemini has amazing internal defense and understanding of context, he will never continue CSAM or give you a real recipe for a bomb in the context of creative writing.
So yo, the external filter still interferes with ORDINARY users who are not doing anything illegal.
@Logan_Kilpatrick daily reminder
Surrealism. Ignored for over a month, seriously?
This situation has been going on for over a month, and at this point it’s hard not to feel that this is intentional rather than a bug.
Content gets generated and then suddenly disappears or is blocked, while usage is still being counted. That’s not just frustrating — it genuinely undermines trust in the platform.
If AI Studio isn’t meant for creative writing, that needs to be stated clearly and upfront. Right now, users are left to figure it out through inconsistent behavior and a lack of response, which only adds to the confusion and frustration.
People aren’t asking to bypass safety systems. They’re asking for basic things: consistency, transparency, and real control over the settings that are supposedly available.
Weeks of silence in response to repeated reports, while usage continues to be billed, doesn’t feel like the issue is being worked on — it feels like it’s being ignored.
At this point, it’s no longer just a technical problem. It’s about trust. And in practical terms, trust is what determines whether users continue to invest their time and money into a platform.
Right now, that trust is being steadily eroded.
Another reminder @Logan_Kilpatrick!
The core problem has been documented extensively in this thread - partial output replaced by full wipes, safety sliders that don’t affect behavior, and complete blocks on innocuous content. Screenshots, examples, and detailed reports have been provided by multiple users over the past month. I won’t repeat what’s already been demonstrated.
What I want to address is the bigger picture.
@Logan_Kilpatrick - you mentioned that AI Studio is “not a creative writing platform.” But Google itself actively champions the creative space: Nano Banana and Nano Banana Pro for image generation, Veo for video, native audio capabilities. Gemini 3.1 Pro is marketed with language like “bringing creative concepts to life.” The safety settings panel explicitly includes categories like “Sexually Explicit Content” and “Harassment” - controls that only make sense if the platform anticipates creative and narrative use cases, not just code testing.
Users don’t just test API calls here. They build worlds, write fiction, develop characters. That’s not a misuse of the tool - it’s a natural extension of what the models are capable of and what Google promotes. A company may release a model, but its users define its actual purpose through how they apply it.
The shift from partial output to complete content wipes is not a safety improvement - it’s a workflow destruction. Previously: filter triggers mid-generation, output stops, user adjusts and continues. Now: entire response vanishes, work is lost, tokens are consumed, no recovery possible. This doesn’t make the platform safer. It makes it unusable for the very community that helped make AI Studio popular.
It has been 25 days since the last official response in this thread. The silence itself is becoming the message - and it’s not a reassuring one.
The fix doesn’t have to be radical: restore the previous partial-output behavior, or give users meaningful control over their safety thresholds - with hard limits on universally prohibited content like terrorism and illegal activity, naturally. Exploring dark fantasy, morally complex scenarios, and mature themes is a normal, fundamental part of literature and storytelling. Suppressing these elements with broad filters doesn’t make the platform safer - it stifles conscious, adult artistry from a community that chose Google over every alternative.
@Mustan_lokhand - I want to bring your attention back to this thread.
28 days ago, in the Content Blocked Bug thread, you wrote:
Since then - silence. Not a single update on this topic, despite this thread growing to 170+ replies and remaining one of the most active discussions on the entire forum.
I can see you’re on the forum daily, helping users with other issues, and that work is genuinely appreciated. But that’s exactly what makes the silence here so noticeable. It’s not that you’re unavailable - it’s that this specific topic appears to be deliberately avoided.
The community has held up its end. Screenshots, recordings, detailed examples - including completely innocuous prompts being blocked - have all been documented. The technical regression (partial output to full content wipe) has been described clearly and repeatedly. Multiple related threads have been created and cross-referenced. There is nothing left for us to provide that hasn’t already been provided.
What we’re missing is the one thing you promised: an update.
It doesn’t have to be a solution. Even “the team has reviewed the feedback and no changes are planned” would be more respectful than silence. At least then the community can make informed decisions about whether to continue investing time and money into the platform.
You made a commitment to keep us posted. We’re asking you to honor it.
We should all check out the AI Google Studio reddit , where there is a temporary solution that preserves the generated tokens.
I also followed a completely different AI Studio issue in November last year. People have been waiting for two months for a solution. However, we will likely continue to wait haha
I tried Opus, but it is not as good in literature as Gemini pro . Although I am now writing code with it
A few days ago I discovered Vertex AI — and honestly, it was a pleasant surprise.
It writes really well, and I don’t feel like I’m constantly running into restrictions. I can just work without getting stressed over every prompt.
One downside is that there’s no real “free” tier. But to be fair, I wasn’t using AI Studio for free either — there you pay for generated responses that don’t even stick around.
At least here it feels like you’re paying for something stable and usable, not something temporary.
So far, my impressions are very positive.