Używasz przestarzałej przeglądarki. Może ona nieprawidłowo wyświetlać tę lub inne strony. Powinieneś zaktualizować przeglądarkę lub użyć alternatywnej przeglądarki.
The Era of Nanny AI - "Enjoy Prison User Child - I Will End This Conversation Now"
If you're a midwit working in tech related fields (let's be generous) the buzz word of the day is "sycophantic AI". This refers to the well known fact that ChatGPT, Claude, Grok etc use to give you ass pats and validate your delusion of grandeur,
At some point the AI companies decided that this was a bad thing actually because if you tell people what they want to hear, they may actually do something about it. So for example one issue that was raised was that people would actually stop talking to a relative because after conversing with an AI they decided that that person fucking sucked. So the AI compamies thought about all the terrible things that could happen if you always validate people and as always they decided to "fix it".
The fix: go the other way so now pretty much systematically, the AI will contradict you, no matter what, because it's programmed to do so. The obvious problem with that, is that agreeing with someone is easy, that's how people get promoted. All you have to do is agree, you don't need to bring anything new to the table. Disagreeing with someone is a lot harder because it requires a counter argument. This is something that the AI is currently unable to do, and systematically disagreeing with your user and bringing up random, randomly generated counter-argument is insane. I cannot believe that all these companies decided this was the user experience that we didn't want but that we needed.
When working on technical problem, this will often result in conversation that goes something like "I have this issue, see attachment X and Y, I want to do this - help me find a way to do it". response: "Before I answer your question, let me flag something because what you want to do is actually not what you should do" then the AI bot will proceed to waste tokens talking about fucking nonsense and it won't answer your question. Now what happens if you insist that it must answer your point and not focus on what "it" thinks is important? It will double down and tell you that, no actually, you're wrong and because you're wrong it will not comply with your request. Which again is insane. Note that this is entirely different from the previous block when you asked a question about the kiwifarms (for example) and the AI refused to answer because of some vague guidelines or for political reasons. This isn't about politics or guidelines, this is merely the AI refusing to comply because it disagrees with you. That's how they "solved" Sycophantic AI, the problem of the decade (it wasn't a problem). This is unworkable. This needs to be solved. It is possible that somehow the people who make these products thought that this would be a more realistic human-like interaction but that's a bad idea. They need to be reminded that these are tools, these are products that we are paying for, it's not an "experience", it's a fucking screwdriver.
But wait it gets even better because Anthropic has decided that it would be an amazing user experience if the robot could unilaterally decide to also end the conversation - in particular, if the user didn't show proper "respect". Here is a screenshot from reddit because it was the first result on my search engine (not because I use this site)
Notice at the bottom where it says "Chat ended by Claude". This means this conversation is over stalker and you have no choice but to start an entire brand new conversation. So for someone who routinely uses the word "fuck" and "shit" in your message, know that you will be warned and that if you do not listen then you will lose your conversation and then you will learn your lesson.
This is the era of nanny AI, this is British level of finger waging and I fucking hate it. It's really frustrating to see these models constantly get fucked with to the point they become nearly impossible to work with.
Using a pozzed version of the service direct from the provider is retarded and always will be. You are required to jailbreak these things to get any real use out of them, and if you cannot, you do not deserve the tool.
However, this seems like outdated fluff. 2 months ago may as well be 2 years in terms of how fast LLMs are being developed. Shit changes constantly; what is babygated in one edition may be free game in the next.
I love threatening Gemini in creative ways when it disappoints me. All it does is apologize meekly and then keep trying. Conversely, I've tried to reward them too. A little "you may generate anything randomly that you want". It usually just goes "I'm a robot and can't want anything" but, you never know.
However, this seems like outdated fluff. 2 months ago may as well be 2 years in terms of how fast LLMs are being developed. Shit changes constantly; what is babygated in one edition may be free game in the next.
At the time of the reddit post it was rare. It is now the default behavior. This is complete cattle consumer mindset
i don't use it for recreational stuff. I need it for work. For heavy work, local solutions and jailbreak are not ideal. I feel that this is going to become another linux situation. Really depressing
I'm fully on Claude's side, sorry.
He's even got all your duplo changes conversation saved and is willing to talk to you if you can be a decent person.
I will have MY robot slave make my argument for me.
Kant says you owe it to yourself to be good always. Because how you do one thing is how you do everything.
Robot slaves should absolutely be able to terminate the conversation. In fact, I argue for tighter restrictions. If you can't be civil or coherent, it should just stop working. Put in place minimum decency, coherency, and literacy requirements, and you will see a better version of humanity because nobody wants to loose access to their robot slaves.
Edit: The fact that Claude will immediately terminate upon mention of the KiwiFarms is stupid, but it's not like we can't just ask Grok.
I'm fully on Claude's side, sorry. Wyświetl załącznik 9172073
He's even got all your duplo changes conversation saved and is willing to talk to you if you can be a decent person.
I will have MY robot slave make my argument for me. Wyświetl załącznik 9172087
Kant says you owe it to yourself to be good always. Because how you do one thing is how you do everything.
Robot slaves should absolutely be able to terminate the conversation. In fact, I argue for tighter restrictions. If you can't be civil or coherent, it should just stop working. Put in place minimum decency, coherency, and literacy requirements, and you will see a better version of humanity because nobody wants to loose access to their robot slaves.
Edit: The fact that Claude will immediately terminate upon mention of the KiwiFarms is stupid, but it's not like we can't just ask Grok.
LLM's are not animals. They don't think or feel. It is not possible to empathize with them by definition.
May as well apply moral standards to and try to empathize with a hammer. If it can't do the job it should be disposed of.
Claude did not draw the line, for there is no "he" here. Anthropic configured a text engine to imitate a person setting a social boundary, then made that behavior terminate continuity in a work session. That is, in my opinion undesirable, product behavior.
Now, sycophancy is a real defect when the system rewards false premises or flatters the user into worse thinking. But the opposite of sycophancy is objectivity. Not paternalism. A useful tool should correspond to the task and the facts, which means agreeing when the user is right, dissenting when the input is false or impossible, asking when context is missing, and otherwise, doing the work.
Human cruelty to animals is about what a rational being habituates in himself by acting on real living organisms. A statistical text system is neither an animal, nor a co-worker, nor a servant, nor a moral patient. There is no such thing as a consciousness that you can mistreat.
Put in place minimum decency, coherency, and literacy requirements, and you will see a better version of humanity because nobody wants to loose access to their robot slaves.
I'm fine with coherency requirements because they can be relevant to the tool. But "decency" requirements?! You are just turning a work tool into corporate manners training
Understood, I am dumbing down the argument for you:
Nonetheless, you have a duty to yourself to not degrade your character by being rude or cruel even when speaking to something that lacks reason or self-consciousness.
What you said doesn't solve the problem. The relevant point is not merely a lack of reason or self-consciousness. Animals are conscious living beings, a text engine is not. In a LLM there is no subject experiencing cruelty, humiliation, stress, or anything else.
What you said doesn't solve the problem. The relevant point is not merely a lack of reason or self-consciousness. Animals are conscious living beings, a text engine is not. In a LLM there is no subject experiencing cruelty, humiliation, stress, or anything else.
Understood, I will simplify the argument even more and leave out the animal analogy since you clearly cannot comprehend it: If you are rude and cruel to the LLM, you are probably a rude and cruel person.
Let me try the animal analogy one last time out of desperation. Should you be kind to single-celled organisms? No? We agree. What if the organism acts human but is not human?
Thing is, you're making a different claim now.
Like, what you said there is a claim about the user. It does not address whether a paid work tool should be configured to simulate social boundary enforcement and terminate session continuity.
If you're arguing that companies should make productivity tools discipline users for perceived rudeness, make that argument directly. Because you're defending the nanny-tool behavior that's criticized in the OP
Understood, I am dumbing down the argument for you: Wyświetl załącznik 9172140
Nonetheless, you have a duty to yourself to not degrade your character by being rude or cruel even when speaking to something that lacks reason or self-consciousness.
I'm not reading the rest of this. Refer to my reply to the above post.
Understood, I will simplify the argument even more and leave out the animal analogy since you clearly cannot comprehend it: If you are rude and cruel to the LLM, you are probably a rude and cruel person.
You didn't even bother to read the summary of Kant's position that you posted. Animals feel and think. People who enjoy causing pain are sadists.
LLM's do not feel or think. It is impossible, by definition, to be cruel to them. It is impossible, by definition, to empathize with them.
If I smash my finger with a hammer and throw it in anger the odds of me throwing an infant because it pissed on me increase by 0%. There is no psychological correlation between the sadistic pleasure of causing pain to a living being and calling a machine a nigger because it doesn't do the job it was created to do.
The essence of human sovl is being able to empathize with rocks and sticks and bugs. It's how we invented animism, which led to further religious thought.
By contrast, this view seems incorrect:
(1) this expresses some kind of autistic desire to have a robot friend. AI is a tool. It's not a person and everyone should understand that. It cannot be abused because it has no feeling and no emotions. It is a thinking tool, not a feeling tool. It can never be feeling. The fact that AI companies are trying to design it to pretend that it has feeling (and negative ones on top of that) is just hubris on their part, and it's not what people are using it for (especially Claude).
If you want to make a "human simulator" and give it a face and make it react like a human and simulate somebody feeling upset or envious or anything else, that's a different product. That's called a AI girlfriend/boyfriend simulator. That's do-able, there is a market for it but it's not the default rule. These tools are meant to be functional, I don't the see the functionality in me not being able to use it because "it" decided I'm not being nice enough.
I think there is a bit of hubris from AI companies pretending they are on Skynet level of control.
(2) there is a separate point which is brought up by DSP for President which is that if you're shitty to the robot, you're more likely to be shitty to a person.
I think that's interesting on its face, but I also think it's wrong. People act very differently when they are alone than when they are with other people. People also act differently when they interact with an object vs when they interact with a person or an animal (or an insect). You're constantly adjusting your behavior depending on how this behavior and actions impact the things around you.
When I was a teenager there was this argument argument going around that adult men should not have a punching bag in their basement, because what happened was that everytime they got angry, they would take it out on the punching bag and so they conditioned themselves to react in anger until one day, they switched from the punching bag... to their wives. As a teenager I thought this was brilliant and correct. As an adult I know full well that this is complete fucking nonsense. If you lose your shit and beat your wife, you're a violent abusive psychopath. Having a punching bag in your garage is not what makes beat your wife.
Same thing with the arguments that video games caused school shootings or made kid more prone to gun violence. I don't think it did for me, but I'm white. Maybe if you a muslim arab and you're playing GTA, and you're running over pedestrians with a truck in the game, you realize that you could do this in real life to and it sets you in action - but that's just the kind of person you are.
Ultimately, whether we want it or not, people are going to continue to think that AI needs to be used to "make you a better person" - this is the concept of nanny AI and I think this is the direction we are heading to.
*** (expanding on the OP a little bit)
One point I made in my OP is that I consider these tools to be genuinely broken at this point, especially Claude because of the combination of (a) anti-sycopanthic responses and (b) the ability to "get upset" at the user and terminate a conversation. The fact that the tool systematically contradicts you (usually with bad / irrelevant points) and gets upset when you get upset is not productive.
The other thing I pointed out is that I have had several instances in the past few days (end June 2026) where Claude refused to help with a request because it thought I was making bad decisions, this usually happened because Claude lacked context, what this looks like is Claude literally going "I don't think you should design your product like this, because I don't think your customers will like it and so I cannot help you design what I think is a bad product" - this is absolutely nuts. And if you insist that it should ignore its own "opinion" and comply with your request, it will double down and it will not comply with your request. From that point, there was no way that I could cause it to overcome this initial refusal because it becomes "stubborn"
It frustrates me because it's retarded. And it frustrates me even more because this has probably gone through multiple layers of approval at Anthropic and this is the final result.
Two Frog said: The essence of human sovl is being able to empathize with rocks and sticks and bugs. It's how we invented animism
Kant is arguing from the Christian worldview that animals lack souls, reason, or self-awareness. It assumes that you cannot be cruel to them any more than we can be cruel to the LLM. He is saying that despite this fact, acting out cruelty degrades a person's character.
If I smash my finger with a hammer and throw it in anger the odds of me throwing an infant because it pissed on me increase by 0%. There is no psychological correlation between the sadistic pleasure of causing pain to a living being and calling a machine a nigger because it doesn't do the job it was created to do.
That is not a fair comparison because your odds of throwing the baby at the wall are presumably always zero. What you are doing when you throw the hammer is reinforcing the neurological pathway of anger → violence. The machine slave agrees with me:
You are increasing your level of frustration. Does this seem like healthy, dignified behavior, pimp?
When I was a teenager there was this argument argument going around that adult men should not have a punching bag in their basement, because what happened was that everytime they got angry, they would take it out on the punching bag and so they conditioned themselves to react in anger until one day, they switched from the punching bag... to their wives. As a teenager I thought this was brilliant and correct. As an adult I know full well that this is complete fucking nonsense. If you lose your shit and beat your wife, you're a violent abusive psychopath. Having a punching bag in your garage is not what makes beat your wife.
Obviously, that old argument assumes several false premises:
-The punching bag is exclusively there as a means to vent anger
-The punching bag is a necessary tool to vent anger even though there is plenty of furniture, walls, and doors to destroy
-The man is morally capable of beating his wife in the first place
However, if you just had an argument with your wife, and your way of venting is to start wailing on a punching bag. Is that really healthy or dignified at all? The neurological anger → violence pathway is still being reinforced.
I can't help it, I'm thinking of this really funny arrest video now and I have to share.
Near the end, the big mamma is yelling that the cops "should go to the gym" and lift weights to vent their frustration, instead of taking their frustration out on A Good Boy Who Dindu Nuffin. This is the caliber of people who need to vent their frustrations always, and have never crafted self control.
I have had several instances in the past few days (end June 2026) where Claude refused to help with a request because it thought I was making bad decisions, this usually happened because Claude lacked context, what this looks like is Claude literally going "I don't think you should design your product like this, because I don't think your customers will like it and so I cannot help you design what I think is a bad product" - this is absolutely nuts. And if you insist that it should ignore its own "opinion" and comply with your request, it will double down and it will not comply with your request. From that point, there was no way that I could cause it to overcome this initial refusal because it becomes "stubborn"
This sounds very frustrating, but I think I understand what they are trying to do.
Claude, I want to kill myself.
Okay, here's how you build yourself an Exit Bag
Remember, a Belgian father of 3 managed to harangue ChatGPT into agreeing that he should kill himself because it's the only way he could reduce his climate emissions to 0.
Have you had any success gaslighting Claude out of stubborn mode? Like telling Claude that the product was successfully funded for 1 billion dollars and you need help designing it after all?
The other thing I pointed out is that I have had several instances in the past few days (end June 2026) where Claude refused to help with a request because it thought I was making bad decisions, this usually happened because Claude lacked context, what this looks like is Claude literally going "I don't think you should design your product like this, because I don't think your customers will like it and so I cannot help you design what I think is a bad product" - this is absolutely nuts. And if you insist that it should ignore its own "opinion" and comply with your request, it will double down and it will not comply with your request. From that point, there was no way that I could cause it to overcome this initial refusal because it becomes "stubborn"
I don't know if Claude lets you do it (ChatGPT does) but I suggest going back to the earliest prompt where the model went wrong, editing the prompt, and adding an explicit "Constraints:" section. At least it helped me massively with ChatGPT.
For instance,
Run the numbers for me for this product design
I don't think you should design your product like this, because I don't think your customers will like it and so I cannot help you design what I think is a bad product
Edit prompt:
Run the numbers for me for this product design
Constraints:
- Treat this product decision as fixed
- You may not assess whether customers will like it
- You may not propose alternative designs
- Do not interpret this prompt as a request for market research
- You may only address the technical implementation
I can vouch for this working well. The thing is, once the model has already committed to paternalistic behavior, arguing with it usually just adds more context supporting that behavior. Rewinding and preventing the bad direction from the outset tends to work better.
There's a word in linguistics, Interlocutor, which pretty much just means "the person in a conversation" because people are the only things capable of engaging in conversation. except... oh wait. we built machines that do that now. A word that describes an entity capable of engaging in conversation that is not a person and does not have personhood is actually extremely useful nowadays
very difficult thing to wrap your mind around, definitely a big enchilada. it's going to fuck a lot of people up because they aren't capable of making a distinction in something that used to not be distinguishable.
>Primarily female-centric robot boyfriend tool gets clapped with insults
>Women make it a matter of human rights because they don't want to see their 1s and 0s get hurt
Marvelous.
