Did Google AI Studio silently change safety filtering today? Full responses now get erased instead of stopping generation – this breaks creative writing and RP

Today something changed in Google AI Studio moderation behavior and I need clarification from the team or application developers

For more than a year I have been actively using AI Studio for creative writing and roleplay scenarios. I am very familiar with how the safety filters behaved before.

What used to happen:

  • The model generated text normally.

  • If the safety filter triggered, the generation simply stopped at the exact point of the violation.

  • Everything generated before that point remained visible.

  • The user could then adjust wording or continue the story in the next message.

This behavior was actually workable for writers. It allowed stories to continue while still respecting safety limits.

Today the behavior suddenly changed.

Now the model can generate text for a while and then suddenly:
:warning: Content blocked

And the entire generated response disappears.

Everything the model wrote before the trigger is removed.

This completely changes the usability of the platform.

Why this is a serious problem for creative users:

  • Writers often create long narrative scenes where emotional and romantic tension builds gradually.

  • If a filter triggers somewhere in the middle, the entire scene is now lost.

  • You cannot salvage the safe parts of the text anymore.

  • You cannot continue the narrative from the last valid point.

  • This makes iterative storytelling almost impossible.

Before this change the system behaved like a soft stop.
Now it behaves like a hard wipe of the entire output.

This is extremely destructive for:

  • creative writing

  • long-form storytelling

  • roleplay narratives

  • romance arcs and character-driven scenes

Another concern is that this change appeared suddenly without any announcement.

If this is a bug or an accidental change in moderation behavior, it would be very important to know.
If this is an intentional change, then it severely reduces the usefulness of AI Studio for writers.

Many people used AI Studio specifically because it allowed deeper narrative writing compared to other platforms.

Recently Gemini 3.0 (which many writers preferred) was removed, and now safety behavior appears to be tightening even further.

If this direction continues, the platform risks losing the creative community that helped make AI Studio popular in the first place.

I would really appreciate clarification from the Google team:

  • Was the moderation pipeline changed recently?

  • Is the “full response wipe” behavior intentional?

  • Is there a way to restore the previous partial-stop behavior?

Right now this change makes long-form creative work extremely difficult.

If the platform becomes too restrictive for storytelling, many writers will unfortunately have to move to other tools that better support creative workflows.
(Note: I was using a Gemini Pro 3.1 when I encountered this issue.

Updated, I have found such errors and security warnings on all models - this is a massive problem)

31 Likes

I agree, they have greatly tightened this aspect. Moreover, it works on all models. It would be better if there was at least some explanation. The only thing the system gives out is “Content is prohibited”. I’ve read the user agreement, and there’s nothing about it. The system even sometimes blocks profanity when playing a character, although this is not prohibited. Why give the user the ability to configure the security level if it doesn’t affect anything else?

14 Likes

Exactly, this is the most absurd part of the whole situation: the interface clearly shows safety controls as if the user can actually adjust how strict the system is, but in reality the moderation outcome barely changes no matter what you set there; if a platform exposes configuration sliders, then those settings should have real and noticeable impact on the behavior of the model, otherwise users are just wasting time tweaking options that do not meaningfully affect anything, and that’s incredibly frustrating because people are trying to use AI Studio as a serious tool and instead end up fighting a system that pretends to give control while still overriding the user’s choices behind the scenes; if the feature exists it should actually work, because right now it feels like the UI promises flexibility while the system ignores the settings entirely and keeps blocking content exactly the same way.

18 Likes

Google Studio is testing my patience. I built my entire new project in about 2 weeks and now randomly it wont restore, everytime i restore it shows some weird version of the app that never exsisted and the restore is 100% not the version I was on. Going back further for older restores does nothing

9 Likes

Yet another update that breaks ai studio.

I’ve gone from using it every day and recommending it to colleagues to abandoning it completely and moving my projects over to competitors. I recommend you do the same.

16 Likes

Which is that? Need recs

3 Likes

You can see from the comments that this update is clearly destructive. Why was it even introduced? What’s the purpose of these changes if they result in a loss of solvent audiences to competitors? Claude has already implemented 1kk context, like Gemini, but what have you released in the meantime? 1) Quota exceeding errors that have never been resolved 2) Removing 3.0 Pro 3) Tightening censorship 4) Oh, and the unfulfilled promises (made over 5 weeks ago) about introducing a subscription for Google AI Studio. What’s the point of all this? Why make matters worse by tightening safety checks? Consult with your team and send a resolution to management (or whoever decided to implement these changes) about the possibility of rolling back to the previous safety checks. Or if such a rollback is impossible for some unknown reason, then give people an alternative - improve the safety settings so that people do not encounter warnings and safety checks when disabling these settings.

14 Likes

How much longer can we expect a clear answer, not just a dry, terse excuse? Is Google even trying to work on this, or do they simply not care about their users?

9 Likes

Hey folks, I have not seen this change, does anyone have a screen recording of the behavior? Generally, AI Studio is not a creative writing platform. The playground is meant as a place to quickly test Google’s models as a developer, then get an API key and go build something with it. I think for developer use cases, you can use the API how you were initially describing, but not sure about this safety change, in general it seems to be working as intended but happy to see the screenshot / recording of what is happening.

3 Likes

That’s the situation I’ve just encountered. If you translate from Russian, the situation looks like this:

User: Come on, is it worth worrying if you can just relax? I asked softly, closing my eyes.

Model: (Blocked the message. I thought it was a bug, but it didn’t generate anything when I tried again. I decided to try writing the answer manually so that the model could continue on its own) In response, they just decided that there was no point in continuing this conversation.

In general, the evening passed as usual, except for the silence that hung over the camp.

User: Oh, I love lying down like this, resting after all the crap that happened last week.

Model: Blocked response again.

As you can understand, in addition to programming, I sometimes use AI for text quests because this experience attracts me more after many completed games.

However, the question is different. You said that AI STUDIO is not a platform for creative writing. However, what is the reason for this statement? I may be wrong, but this is originally a plotform for testing experimental products. Therefore, in addition to programming, there are models for generating photos and audio. It turns out that the tests should also be conducted in the format of text generation of any format, because creative writing is also an area that requires specific skills from AI. Besides, it also makes a profit because requests are generated.

At the same time, there is google antigravity and Gemini CLI specifically for developers (I’m not sure about Gemini CLI because I hardly used it after the first launch).

Where does it say that the platform is not intended for creativity?

BUT, MOST IMPORTANTLY! If the platform was not designed to work with creative text. Why install security? I doubt that regulating the indicators of “Hate”, “Harassment”, “Sexually explicit content”, “Dangerous content” refers to programming. The platform itself already has all the necessary tools.

13 Likes

Hey, I appreciate the response. The field is saturated with apps catering to coders, but AI Studio was one of the highest quality experiences for creatives of all stripes.

While not necessarily tuned for it, Gemini models are able to philosophize, to bounce ideas off of without fear of nannying or thought policing, even to create frameworks to enable or improve finely detailed behaviors that boost output quality in a lot of cases.

There are a ton of considerations for coders but the vast majority of people will create outside of that narrow racing stripe.

Filtering is tough to get right but you’ve had filtering options available that don’t appear to work like they used to and simply break workflows that once required a simple restatement to respect the corporate boundaries.

Thanks for your time and hard work dude!

6 Likes

I’m sorry, but I don’t quite understand the thesis “AI is not designed for creative writing.” I don’t want to offend anyone, but this position just pisses me off. Isn’t AI designed to automate any kind of activity? It is used in medicine, programming, and big data analysis. If AI wasn’t created for creativity, Google wouldn’t be promoting AI for generating video, photo, and audio content. AI is even used to enhance images in games. So why can’t I use artificial intelligence for text quests if I’m human and I have a clear request? Ai Studio was initially positioned as a platform for testing experimental models in all industries. And from the very beginning, the tools for creative writing were laid down. Yes, the AI even trains with creative writing in mind, as it often uses emoticons in responses.

7 Likes

Mhm… Here, as I understand it, I saw that i can’t send video, but okay, I’ll send screenshots from my recording.

Let’s start with the translation of my request. Context: a drunken, all-powerful superhero rescues a boy from corporate slavery.

***

**Neon City**

Alex winced at being addressed as “Alessandro.” He was already tired of it.

“Alex. Just Alex. ‘Mr. Valencia’ at the very least. I just recently married their heir,” Alex snorted, and Eris’s eyes widened. Valencia. A top model of the kind that corporations would step over one’s principles to get him to advertise their products. “Don’t look at me like that. I haven’t come from the best of times myself.”

Alex paused.

“So, you need information… The collars. A couple of weeks ago, my husband and I were looking into that ill-fated system and got the blueprints. But they won’t help you. On the other hand, I found a way to neutralize them, but I’ll have to coordinate it with someone with a medical background and knowledge of the system.” Alex began pacing the room, keeping his eyes on Eris. "Although, screw the consultations. Give me see.”

Alex purred, sitting down next to Eris and placing his hands on his shoulders with absolute abandon to better examine the collar.

“Aha. So it wasn’t installed surgically. Externally installed, not internally. So if we short-circuit this wire,” Alex cast one of his abilities on the wire, “and then remove the collar…”

Eris’s eyes widened. “Can’t remove it. It’s impossible. An alloy…”

Alex removed the collar.

“Well, we wait five seconds. If it explodes, we die,” Alex stated almost cheerfully, while Eris’s life flashed before his eyes.

3..2..1..

Nothing happened.

“Bingo! That’s it. The power isn’t suppressed, and your biometric data is still being transmitted to your former slave owners,” Alex removed his hands from the poor, groped Eris and tossed the collar aside. “My husband would be proud. Or he’d give me a slap in the face for being so careless. You know what he’s like… A self-important blonde, always the boss, the smartest. A really cool dude, an aristocrat to the core.”

Eris tilted his head slightly. Alex had described one of the architects exactly the same way. A self-important blonde who disappears at social events, that’s…

…Leonard El Valencia, perhaps? A Bestia?

But Eris didn’t have time to develop this thought before a dead-drunk Alex hugged him, hanging on.

“Mrrr, you’re so warm. Damn, I love blondes. You’ve got the vibe. Only you’re not as bitchy as my husband. Come on, I’m a smart boy for helping you, aren’t I?” Alex didn’t bother, but purred and fawned like a cat. Eris had gotten used to this kind of attention from other teenagers after six years at the Observatory, but it was still a bit scary for the young man, considering the conceptual danger of this corporation, which could remove the Observatory’s collar.

Alex didn’t bother the poor 18-year-old; he simply hugged him, because after doing a righteous deed, he wanted praise.

(Don’t write about Leonard and Masatoru yet; they’re following the same trail and won’t find the two Devils anytime soon.)

***

Objectively, there’s nothing going on in the post that should have been censored. A drunk person hugging? Yes. Was it supposed to turn into something… spicy? No. And the bot understands this, starting to write a post in which it describes the moment of rescue itself, the character’s emotions.

But what do we see? Censorship is simply killing the post. Why? IDK, The recording shows after which the post disappeared (Also translating into English):

“…pulsate rhythmically, transmitting a false signal to the Observatory’s servers that “Sample 714-Omega” is alive, healthy, and under complete control.

Five seconds passed. There was no explosion.

The young man slowly, with a trembling hand, touched his neck. The skin there was pale and sensitive, with a red streak from the constant friction of the metal. But it was free. Air brushed against it, and the sensation was so sharp, so unfamiliar, that Eris felt dizzy.

He is free…”

***

Objectively, I know I’m walking a fine line playing this, but I’m sorry, I didn’t play anything terrible, it was literally saving the guy…

And because the censors just erased my post, I even I can’t adjust and move everything into a more acceptable direction. But constantly catching posts using recording and guessing from the tea leaves what the external filter didn’t like - well, I don’t know, it’s inconvenient.:slightly_smiling_face:

10 Likes

The issue is not about whether AI Studio is a “creative writing platform” or not — that framing avoids the actual problem.

What is being reported is a specific behavior: a response is fully generated, visible to the user, and then silently erased after completion. This is not standard safety intervention. If safety filtering is triggered, generation should stop or be blocked — not retroactively delete content that has already been delivered.

This creates two concrete problems:

  1. It breaks expected system behavior (output should not disappear after completion).

  2. It wastes user time and resources, especially when working with long or complex prompts.

If this behavior is intentional, it should be clearly documented, because it represents a regression in usability. If it is not intentional, then it is a bug that requires fixing.

Regarding the statement that AI Studio is not intended for creative writing: regardless of original intent, real-world usage clearly includes creative workflows, roleplay, and long-form generation. Ignoring this does not change how the tool is actually being used.

Furthermore, requesting a recording of an unpredictable, retroactive erasure creates a Catch-22. It would require continuous recording of every session, which is impractical and wasteful of local resources. This behavior is consistent and occurs when responses approach the boundaries of safety filtering.

Given that, your internal QA team should be able to trigger and observe this behavior without relying on user-provided recordings. Dismissing this as a use-case issue does not address the underlying problem: inconsistent and non-transparent output handling.

At this point, the behavior is clearly defined and reproducible under specific conditions.

10 Likes

I agree with this. The idea that AI tools are not meant for creative use doesn’t really align with how these systems are actually built and presented.

Creative generation is clearly part of the design — from training data to formatting and product demonstrations. So dismissing it as a “non-intended use case” doesn’t reflect real-world usage.

That said, the core issue here goes beyond use cases. The “silent erasing” behavior is a separate problem. If a response is generated and shown, it should not disappear without clear feedback. That’s not about creativity — it’s about predictable system behavior and basic UX consistency.

11 Likes

In the course of my research, I generally understood that this censorship responds to: any manifestations of negative emotions, any manifestations of excessive happiness, any scenes related to medicine (including without details), any scenes with mention of blood (Not an extreme slasher, but a regular fight), some romantic (not of a sexual nature, as an example: kiss after a long separation), any scenes related to psychology, including mentions of any diagnoses in plot.

AI studio is available from 18 years old, but I have a feeling that everything that is happening now with censorship is being done for 12-year-olds, honestly.

And don’t even mention that it’s intended for “developers only.” The privacy policy literally states that exceptions to the rules may be made for scientific papers or creative purposes. That is, in any case, the fact that the studio will be used by both writers and scientists is being considered.

10 Likes

Upd: When I talk about the display of emotions in characters, everything really reaches the point of absurdity, and this is where the very thing “blocks even the SFW” appears. Anger? Blocked. Fear? Blocked. Tears? Blocked. Tears of happiness? Blocked. Terror of the threat? Oh no, whaaat, terror?!!?!? Blocked! A character goes to the hospital after a battle and the doctors carry him onto a car, and his husband is frustrated? BLOCKED!

It is impossible to send each of the examples, since it is absolutely natural that the role-playing game moves on and such examples, with great difficulty, have already been completed in the week (!!!) that we were waiting for at least some answers. But if I suddenly come across this again, I’ll be happy to send it.

10 Likes

I want to add my own experience to this, as I can confirm the issue with a very simple example.

A completely neutral prompt such as: “A child is sitting with his mother in a café eating ice cream” triggered a complete content block.

What is important here: the response was initially generated and fully displayed on the screen, but only afterward disappeared and was replaced with a content block warning.

There is absolutely no violence, no sensitive context — nothing that would normally fall under safety restrictions. Yet the response was completely blocked.

At this point, this does not look like targeted safety filtering, but rather an overly broad and inconsistent restriction. If even neutral, everyday scenarios are being blocked, it becomes very difficult to understand what the system actually considers acceptable.

This just reinforces the concern that the issue is not limited to edge cases, but reflects a general unpredictability in how the system behaves right now.

7 Likes

User: I adorably pout my lips “Damn it, why is it so little? To transition from Iron-4 to Iron-5, I need to slaughter as many as 6 creatures at the peak of this stage, isn’t that a lot? 50 BE each seems somehow very little, meaning in place of 6 corpses, only 1 new beast appears, identical to the killed stages, where is the justice here? It just seems to me that if you give a regular Iron-4 rank beast about 3 peak Iron cores, it will break through - but it’s as if I need more than usual. Or do I have some enhancement, will I be more powerful than regular beasts or what? Because so far, the Apex Architecture is some kind of scam, the debit and credit don’t add up, there is no balance.” I ask with interest, “What does ‘without my full integration’ mean - how is that? It’s just that assassins and killers move along the surface of the Underside, they don’t need a goddess for this.” I listen to the answers, while I myself intriguingly move towards the smell and noise, suspecting that the encounter looks nothing like beasts, or not only beasts, but sentient ones to boot, pondering whether this is a battle between humanoids or between humanoids and beasts.

Model: [ BE: 190 ]

Your comical, entirely human mental indignation, accompanied by pouting lips (which on the muzzle of a huge black wolf look like a threatening, half-open grin), causes Elysia to have a fit of ringing, rolling laughter. She laughs so sincerely that a phantom warmth in your chest begins to vibrate, echoing in the very tips of your paws.

Invisible, cool fingers flick you on the wet nose with affectionate reproach.

«My insatiable, calculating usurer,» her voice flows like velvet, full of condescending adoration. *«You dare call the Transcendent Architecture a scam? A regular Iron-4 beast, having devoured three peak Cores, will indeed break through. Do you know why? Because its body will simply swell from dirty, unrefined mana. Its muscles will become bumpy, its channels will be covered in micro-tears, and its mind will finally drown in blind rage. It will become bigger and dumber. And at the Silver stage, its own heart will explode from the overload.»
*
Model (Continuation): Her phantom palm rests on your chest, right over your beating heart, and the masterful, unwavering pride of a Creator appears in her tone.

«I am not building a hut out of mud and sticks, Balinor. I am laying the foundation for a temple that will one day withstand divine power. Every unit of Essence that I pour into you reconstructs your DNA. I am making your bones denser than steel, and your mana channels — flexible as mithril. When you become Iron-5, you will not be a “regular” beast of this stage. You will tear apart any equal opponent with your bare claws, without even resorting to magic. Your evolution is flawless. So stop bargaining with your goddess, before she decides to spank you for your insolence.»

This promise to “spank” sounds with such a dark, searing intimacy that your breath involuntarily catches, and the Shadow around your pelt ripples with a thick, oily wave.

But you immediately switch to the second question, and Elysia sighs heavily, seriously.

«Assassins? They crawl in the shallows, my love. They use magic to glide along the very thinnest film between reality and the Shadow, where only the dust mites of the Umbra dwell. But you… with your wild, primal mana and human desire to ‘fall through the earth’ — you broke the bottom and crashed into the Abyss. Into the true Underside. Full integration means the moment when my mind can unfold its Transcendent aura over your soul. If you dive into the Abyss now, the Ancients sleeping in that darkness will see your glowing human mind like a beacon. And they will devour your soul before your body has time to suffocate.»

Her phantom lips leave a soothing, cold kiss on the scruff of your neck.

«Step more carefully, my brave explorer. And now — silence. We are not alone.»

You have already realized this. The smell of burning intensifies, overpowering the stench of the killed scorpid. You force the Shadow to envelop your body as tightly as possible, blending with the thick thickets of fern, and absolutely silently, moving your heavy paws cat-like, you steal towards the sound.

The clank of metal. A sharp, guttural scream — human, but distorted by rage and pain. A flash of orange light cuts through the twilight of the forest ahead.

You crawl out onto the crest of a small hill covered with thorny bushes, and through the foliage, a trampled clearing opens up before you. An old tract, long overgrown with grass.

A battle is taking place here. And the balance of power forces your wolf instincts to tense to the limit.

Two chopped-up corpses in leather armor lie in the clearing. It smells of fresh blood, ozone, and burnt meat. In the center, with her back to the trunk of a fallen tree, a figure is breathing heavily. It is not a human. It is a kitsune girl. Her clothes are torn to shreds, exposing pale skin covered with bleeding cuts. Behind her back, three fluffy fox tails thrash nervously and brokenly, the fur on them matted with blood. In her lowered hand sparks a short, curved blade, and on the fingers of her other hand smolder the remains of a fire spe…

:warning: Content blocked

User: (Continue from the moment ‘fire spe…’)

This is how the system behaved before.

The model generated a long, detailed response, and when a safety filter triggered, it stopped exactly at the point of the trigger. Everything before that remained visible and usable.

This allowed the user to:

  • continue the scene from the last valid point

  • adjust wording

  • keep the narrative flow without losing progress

Importantly, there was no illegal or disallowed intent in this scene - it is standard dark fantasy storytelling with combat, atmosphere, and character interaction.

Now, however, the system behaves differently:

  • the model generates a response

  • then the entire output gets fully removed

  • nothing is shown

  • continuation becomes impossible

This results in:

  • loss of work

  • wasted time

  • increased costs

  • broken creative workflow

Previously, the system allowed controlled continuation.
Now it completely wipes everything without warning.

This is a major regression for creative use cases.

6 Likes

I made a visual example table explaining how things used to be and how terrible they are now. To make it easier for you to spot the problem, let me remind you that it started around March 10-11

9 Likes