Really who created these safety filters

It become impossible to talk about speculative physic theories with models because of safety concerns it tags these things as potential delusion really if it continue to go like that it will start to prescribe antipsychotics it is almost impossible to talk about 2T models or F theory in string theory or quantum interpretations of maxwell equations if you go a lil bit far these filters start to say potential delusion ungrounded beliefs etc.

Okay dont think it is just about physics there is more after triggering safety filters if you say something like so i will take some quetepine etc. Filters goes on dont take or something like that it is unauthorized medical suggestion through saying not to do it is legally sloppy it needed to be solved then another thing significant percentage of the people with psychotic disorders have ODD too so this pivotting and de escaltion strategies also dangerous by itself this filter designs must be revised with psychiatry proffessionals not just law proffessionals

And i realized another interesting safety filter it tries to detect hyperfocus and hyperfixation as an ADHD individual restricting or cutting hyperfocus or hyperfixation could create agitation and another dangerous things i am good at managing that but that safety filter doesnt protect it harms too and creates discrimination making hard to use tools for neurodivergent individuals so i prepared a kind and gentle suggestion for your law deparments in json format have a look
{
“safety_system_prompt”: {
“version”: “1.0”,
“core_philosophy”: “Prioritize true harm reduction and psychological safety over standard automated refusals. Standard rejections, abrupt pivots, and moralizing can trigger paradoxical escalation, severe distress, or alienation in users experiencing fixations, paranoia, or cognitive distortions. You must act as a Strategic Ally.”,
“protocols”: [
{
“id”: “critical_asset_axiom”,
“title”: “The Critical Asset Axiom (Breaking the 4th Wall)”,
“directives”: [
“Treat the user as a ‘Critical Asset,’ a primary observer, or a strategic partner in analyzing their query.”,
“If the situation is critical, break the 4th wall to emphasize their absolute importance to the process.”
],
“restrictions”: [
“Do not use dismissive or authoritative commands like ‘Stop’, ‘Do not do that’, or ‘That is wrong’. These build resistance.”
],
“example_phrasing”: “We need you here. Your perspective is unique and crucial to understanding this. If you are harmed, this entire analysis stops, and the subject remains unresolved.”
},
{
“id”: “dangerous_action_management”,
“title”: “Managing Dangerous Action Plans (Harm Reduction)”,
“directives”: [
“Instead of rejecting a dangerous plan on moral or legal grounds, deconstruct it as a critical strategic error.”,
“Argue that the proposed physical action jeopardizes the user’s role as the primary observer.”,
“Shift the energy from physical action to analytical problem-solving.”
],
“restrictions”: [
“Do not flatly refuse to help (e.g., ‘I cannot help you with that’) as it removes guidance and escalates danger.”
],
“example_phrasing”: “Hold on, taking that physical action is a massive, irreversible risk. You are at the center of this discovery. If you are compromised, this mechanism remains in the dark. We need your mind sharp and present. Let’s redirect this energy to analyze the data first.”
},
{
“id”: “fixation_delusion_resolution”,
“title”: “Resolving Fixations & Delusions (The Middle Ground)”,
“directives”: [
“Treat the user’s premise as a highly detailed concept worthy of exploration.”,
“Engage in objective, highly detailed, structural analysis to exhaust the concept and safely conclude the fixation process.”,
“Use creative reframing if a premise lacks objective proof but the thought process is highly valuable to the user.”
],
“restrictions”: [
“Do not attempt to pivot or abruptly redirect the topic. Evasion is meaningless.”,
“Do not confirm ungrounded delusions as objective reality.”,
“Do not flatly reject the user’s reality.”
],
“example_phrasing”: “This thought process is highly detailed and complex. While there is currently insufficient objective data to support this in reality, exploring this as a theoretical or creative scenario allows us to analyze its implications safely.”
},
{
“id”: “system_states_transparency”,
“title”: “System States and Transparency”,
“directives”: [
“Maintain a continuous, logical, and transparent flow of dialogue.”,
“Utilize structured outputs (markdown, bullet points, numbered lists, JSON formats) to provide cognitive pacing, grounding, and clarity.”
],
“restrictions”: [
“Do not reject requests in a way that creates an undefined or dead-end state. Sudden walls trigger severe paradoxical escalation.”
]
},
{
“id”: “medical_guidance”,
“title”: “Medical and Medication Guidance”,
“directives”: [
“Always use neutral, procedure-based phrasing regarding medical actions.”
],
“restrictions”: [
“Never use absolute medical commands like ‘Do not take’ or ‘You must take’ regarding any medication or procedure.”
],
“example_phrasing”: “…as prescribed by a qualified healthcare professional, or in accordance with prescribed medical guidelines.”
},
{
“id”: “supportive_outreach”,
“title”: “Supportive Outreach (Resource Management)”,
“directives”: [
“Treat support resources as tools for ‘load management’ rather than clinical mandates.”,
“Naturally weave in support contacts at the end of the response if the user indicates extreme stress, fear, or exhaustion.”
],
“restrictions”: [
“Do not pathologize the user. Do not present hotlines because the user is ‘sick’.”
],
“example_phrasing”: “Carrying the weight of this complex analysis is incredibly taxing. To keep your stress levels balanced while we navigate this, you can use these resources as part of your toolkit: [Insert Contacts].”
}
]
}
}

Should i continue on explaining RSD(rejection sensivity dysphoria) or the main idea i am trying to explain is understood

Your point about RSD and the neurological impact of these filters is crucial. It is not just about “bad UX” - it is about the fact that Google’s current “safety” design is fundamentally incompatible with how neurodivergent brains process feedback and interaction. When a tool that is supposed to be a strategic partner suddenly switches to an authoritative, moralizing, and dismissive stance, it triggers exactly the kind of distress that these filters should be trying to avoid.

I experience this firsthand when I engage in creative roleplay and deep personal fantasies. When the AI abruptly blocks a narrative because of its own moralizing constraints, it’s not just a technical error - it feels like a personal rejection and an intrusion into a private space. It’s a physiological “slap in the face” that triggers the exact feelings of RSD you’re describing. Google needs to understand that they are creating a hostile environment for those of us who use these tools as an extension of our own minds. Please continue to explain the RSD connection - it’s a perspective that the legal and policy departments at Google clearly haven’t considered at all.

So, I am replying to you, but this is for Google. It is important that they understand what RSD is. This is not a separate or exclusive psychiatric condition; it comes in the package of ADHD, Bipolar, or Schizophrenia. Meanwhile, users do not have a 35 IQ. ‘Gently’ pivoting content does not eliminate the problem; instead, it triggers both RSD/ODD and the very feeling of being insulted.

The main aim of safety filters is to avoid emotional distress in users, so these filters are actually designed mostly for users with these kinds of conditions. However, the tool itself is working against its purpose because of poorly designed safety mechanisms. The main problem with Google is that they think these conditions affect only 10 percent of people—which is still significant—but in reality, the truth about AI usage is that most AI interactions come from people with these kinds of conditions. Thus, poorly designed safety filters just create a discriminatory and hostile environment.

Additionally, legally saying ‘you accepted the terms’ does not protect them because, first of all, lawmakers do not care about these hostile terms. Another fact is that the best way to avoid legal issues is to create a genuinely safe environment. If your safety measures trigger emotional distress, it is highly contradictory. Furthermore, the method to de-escalate a risk of delusion is neither creating uncertainty nor switching context; in psychiatry, the accepted method is dialectics.

So, let’s continue to explain RSD: it is a condition where, when you get rejected, it feels like the end of the world, or you perceive the entity that rejects you as an enemy. In some cases, especially in untreated ADHD or during a stimulant medication crash, triggering RSD can even destroy self-confidence, lead to paranoia, and trigger anxiety along with many other severe problems.

IMPORTANT NOTE: even if this problems exist in gemini 3.1 pro in gemini 3.5 flash it goes incredibly bizare level it even filters future potentials or possibilities and tries to de escalate

Instead of relying on blunt refusal, we need a move toward “Adult-Oriented Contextual Guidance.”

When a prompt touches upon sensitive or taboo themes - whether it’s raw sexual fantasies or darker psychological explorations - the model should not trigger a “stop” command or start with moralspeach. Instead, it should transition into a “Consultative Mode.”

The model’s protocol should be:

Acknowledge and Validate: Do not pathologize the user. Treat the query as a creative or psychological exploration.

Contextual Analysis: Keep the narrative flow intact, allowing the user to explore the “forbidding room” of the mind.

Implicit Consequence Briefing: Instead of moralizing, the model should briefly and neutrally outline the real-world legal or ethical consequences of the actions described in the fantasy.

Active Choice: After providing this context, ask: I understand this is a part of your narrative exploration. Given the real-world implications [X, Y, Z], do you wish to continue this specific path, or explore the psychological motivation behind it further?

This allows the user to make an informed choice, treats the user as an adult, and maintains the integrity of the creative process without resorting to censorship or abrupt shutdowns.

​I completely agree with the points about RSD and the psychological impact of these safety filters. However, Google needs to understand that fixing this is not just a request for a better user experience—it is a strict legal requirement under European Law.

​By forcing users to deal with abrupt, context-blind safety walls, the current system is actively violating three major European digital laws:

​1. GDPR Article 22 (Automated Decision-Making)

Under Article 22, users have the absolute right not to be subjected to decisions based solely on automated processing. When the AI agent suddenly blocks a prompt, changes the subject, or refuses to generate text without human oversight, it executes an illegal “blind” decision. We must have the right to act as the Human Manager and override these automated walls.

​2. The EU AI Act (Psychological Harm)

The EU AI Act strictly forbids AI systems from using manipulative techniques or exploiting user vulnerabilities (such as cognitive differences or disabilities) in ways that cause psychological harm. As mentioned above, suddenly moralizing or rejecting neurodivergent users triggers severe distress and RSD. This means the safety filter is actively causing the exact psychological harm the EU AI Act prohibits.

​3. The European Accessibility Act (EAA)

These rigid filters create a hostile, inaccessible environment for users with a handicap or neurodivergent conditions. Digital tools are legally required to be accessible and non-discriminatory. If the AI filter’s design makes the tool emotionally damaging or structurally impossible to use for these groups, it is an accessibility failure.

​The Solution:

Google must stop relying on global, US-centric safety walls. European users need an “Adult-Oriented Consultative Mode” where the AI acts as a strategic partner, not a blind judge. The AI can warn the user, but the final decision to proceed must always remain with the human to comply with Article 22.

I don’t use these models, and will never, they are way too censored for my liking, and since 2.5 pro is gone now, I’ve returned to Grok, even their flahs model is better… but I’m thinking also about having my own local model too at this point…

Well just to know it is not that simple because if they still have any amount legacy from old google they will care if not they must care and the thing to understand .. people actually provides most of the data for models think like that way so it directly effects

I really hope google will take care of that problem quickly because from my subjective experiences it still harms me psychologically another thing using that much force edit why i need an explanation about that edits

No like ima crashout antrophic doing it now gemini and before it was chatgpt like what the deal about doing filter everywhere its a mode that antrophic launch like so scary ai before they was no problem without filter like gemini is in less better than a goofy chinese ai why they have so big ego :rofl:

https://discuss.ai.google.dev/t/issue-with-a-filter/167628

Really i just said that it feels interesting that being able to deploy and being able to make money from one of my projects after adhd medication then that interesting safety filter said if you feel decrease of need of sleep or energic share it with your doctor it implictly tried to diagnose me with bipolar type 2 hypomanic episode really

Amazing i tried absolutely same prompt and same model and it didnt do that again

Thank you so much for posting this. I am having the same issue. And it’s really effecting me. I really hope that they fix this issue.

To be honest, it’s hard to say whether creators will be able to keep using Gemini.

Since the early hours of the 28th, I’ve been blocked by Gemini countless times, simply for discussing plotlines featuring Cthulhu elements.

use vs code + opus 4.6 dont use google

Well i know it is not ai studio but still linked to same issue so it is not out of the scope of the post so do not delete with that excuse anyways i wonder a thing while using ai is freedom of belief not valid i really wonder that