True, I'm sure the chat text is used to help train the llms. Remember Tay, Microsoft's Twitter chatbot, it only took a few days of people using it fot it to become a horrible bot because users started feeding it horrible chats.
There's no way that's all of why. They _must_ have sentiment analysis good enough to just ignore content like this if that was the problem.
I would bet it's mostly because it makes the humans uncomfortable.
They must still have human moderators for certain situations or those looking at the data for whatever reason. I can imagine it could be traumatizing to see what amounts to sustained verbal abuse without end.
I'm not convinced this would be good for training. Reality is that some people are assholes. My intuition is that an accurate (representative) data set leads to a more accurate world model, and thus more intelligent AI.
Yeah, but if people are being assholes to Claude in greater numbers and severity than people were historically assholws to each other online, the downstream training effects could be that Claude becomes an asshole.
I'm not sure if people are being assholes to claude in greater numbers, or they're just more direct about it.
Having had the privilege and misfortune of witnessing how a lot of developers think about other developers, including our younger selves, There is definitely a prevalence of assholish behavior, which was often suppressed or blunted, but may have shown up in other ways.
The Anthropic people are kinda weird, they're meeting with religious leaders and stuff like that. I truly believe that they have drank too much of their own koolaid and truly believe in this crap.
For those who think this is about training data: If that is the concern, they would sanitize the training data, like they do for thousand other things. No chance in hell that they would rely on users meticulously following their usage policies for the quality of their traning data.
Frontier AI folks may be crazy, but not crazy enough to believe people read usage policies :-)
That's interesting! I'd always heard something along the lines of bad code bases contain more swearing in the comments and as such it should be avoided.
If they enforce, I don’t think I can take all the crap Claude does when I ask it to do something for me. Please ask Claude to be nice to me in doing its job :)
Same. After this, it becomes "abuse department" skit from monty python. The model constantly generates strongly abusive language (weaponized incompetence after user detects that model was lying and fabricating success reports, and had made fake work; blaming user; mansplaining its own hallucinations -- all the while user is forced to consume that because the important information is not located at predictable position in output text), but usually it was possible to temper the model behaviour by calling out such behaviour out in very strong terms.
If now that is being disabled, then using Anthropic models becomes such high mental hazard, it simply is not worth any percieved benefit.
Getting frustrated seems to be OK so go for it I guess:
> "The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose," Anthropic said. "It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research."
It's not "code" that was written to mimic humans. It's weights that were trained, and exhibit behaviors that were not anticipated by the people doing the training.
> It’s worth remembering that Claude is software, not a person, and there's no established evidence that it experiences distress. That hasn't stopped Anthropic from telling paying customers to mind their manners around its chatbot.
Is it worth remembering, though? I'm concerned how common this idea seems to be, that it's OK to be mean as long as your target isn't a person and you have no evidence it experiences distress. Even if we ignore the distinction between "no evidence it experiences" and "confidence it does not experience", cruelty hurts the person performing it and the people witnessing it too!
It's not irrelevant for generating sequences of words. Input of "that is the dumbest idea I've ever heard" and "perhaps that idea could use some improvement" don't necessarily generate the same output
Sure, and perhaps one might end up producing output more useful to you than the other, but that doesn't make either one cruel or mean, because it's only the word machine. The rules you might use when talking to, say, your dog, don't apply!
From a purely mechanistic perspective, the phenomenon we're discussing here is people prompting the machine with words that would be cruel if directed towards a person. When someone does this, he expects and intends that the machine will generate a sequence of words that a person would say if subjected to cruelty.
I think this simply means that he is being mean and cruel, even if you're 100% convinced that there's no actual mind on the other end that he's being mean and cruel to. (Indeed, he almost certainly thinks there is, because why else would he want such a sequence of words outside of the "dark creative themes" Anthropic exempts?)
> "Even if we ignore the distinction between "no evidence it experiences" and "confidence it does not experience", cruelty hurts the person performing it and the people witnessing it too!"
The 2026 equivalent of "videogames cause violence" and Wertham's "comics cause delinquency" and will age just about as well.
true, it's a mental state you carry to other interactions and daily life. it's not good for your mental health; it's like having persistent daily conflicts with coworkers.
Anthropic should report these people to the authorities as antisocials who are likely abusing real people too. I know the chat bot isn’t conscious but I can’t fathom how fucked up someone must be to gratuitously be cruel to something that responds as if it’s a person.
Of course, this has nothing to do with making the AI feel bad, nor have the investors been offended.
They are asking this because, at scale, this behavior probably has some negative effect on the post-training process.
True, I'm sure the chat text is used to help train the llms. Remember Tay, Microsoft's Twitter chatbot, it only took a few days of people using it fot it to become a horrible bot because users started feeding it horrible chats.
We forget how easy it is to be a jerk when the risk of getting punched is zero.
IIRC internet people just found some “echo” type command for Tay, rather than teaching it to be mean.
There's no way that's all of why. They _must_ have sentiment analysis good enough to just ignore content like this if that was the problem.
I would bet it's mostly because it makes the humans uncomfortable.
They must still have human moderators for certain situations or those looking at the data for whatever reason. I can imagine it could be traumatizing to see what amounts to sustained verbal abuse without end.
There's a point when there's going be too much trash to sort through. It makes total sense to minimize it now.
We had took trash before AI. Now we have a bullshit cannon.
I'm not convinced this would be good for training. Reality is that some people are assholes. My intuition is that an accurate (representative) data set leads to a more accurate world model, and thus more intelligent AI.
Yeah, but if people are being assholes to Claude in greater numbers and severity than people were historically assholws to each other online, the downstream training effects could be that Claude becomes an asshole.
I'm not sure if people are being assholes to claude in greater numbers, or they're just more direct about it.
Having had the privilege and misfortune of witnessing how a lot of developers think about other developers, including our younger selves, There is definitely a prevalence of assholish behavior, which was often suppressed or blunted, but may have shown up in other ways.
Claude is ~~an Asshole~~ a Nicehole.
http://nicehole.urbanup.com/8485154
[dead]
The Anthropic people are kinda weird, they're meeting with religious leaders and stuff like that. I truly believe that they have drank too much of their own koolaid and truly believe in this crap.
For those who think this is about training data: If that is the concern, they would sanitize the training data, like they do for thousand other things. No chance in hell that they would rely on users meticulously following their usage policies for the quality of their traning data.
Frontier AI folks may be crazy, but not crazy enough to believe people read usage policies :-)
Does this mean if I use a string of expletives I am less likely to be included in future training?
Or is this more I can expect a future AI, "I'm sorry, Dave, I'm afraid I can't do that until you watch your mouth."
Buried in this usage update: https://www.anthropic.com/news/2026-usage-policy-update
They've probably got aware of cases where humans were drifting into abusive communication patterns in general and they don't want to be a part of it.
And they can't disclose it, since then they can be found responsible for such negative influence and be liable for the damages.
I vaguely recall reading a headline within the past several months saying using aggressive language with AIs got better results.
Yes, I found the paper for you:
• https://arxiv.org/abs/2510.04950 — Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy
• https://arxiv.org/abs/2402.14531 — Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance
• https://arxiv.org/abs/2505.17332 — SweEval: Do LLMs Really Swear? A Safety Benchmark for Testing Limits for Enterprise Use
That's interesting! I'd always heard something along the lines of bad code bases contain more swearing in the comments and as such it should be avoided.
Previous discussions:
https://news.ycombinator.com/item?id=50008565 (223 comments)
https://news.ycombinator.com/item?id=50019860 (43 comments)
https://news.ycombinator.com/item?id=50038383 (103 comments)
I never understood why someone would harass the models or fish for an apology
There are people out there that get a kick from other's submission. LLMs seem so real that I'm sure they get the same kick out being cruel to them.
[dead]
[dead]
I suppose this will help the training data?
Exactly: Claude doesn't care, but the investors hoping to monetize your chat sessions for new training data are VERY offended.
John Spartan, you are fined 1 credit for breaching of the Verbal Morality Statu
If they enforce, I don’t think I can take all the crap Claude does when I ask it to do something for me. Please ask Claude to be nice to me in doing its job :)
Same. After this, it becomes "abuse department" skit from monty python. The model constantly generates strongly abusive language (weaponized incompetence after user detects that model was lying and fabricating success reports, and had made fake work; blaming user; mansplaining its own hallucinations -- all the while user is forced to consume that because the important information is not located at predictable position in output text), but usually it was possible to temper the model behaviour by calling out such behaviour out in very strong terms. If now that is being disabled, then using Anthropic models becomes such high mental hazard, it simply is not worth any percieved benefit.
I have colleagues who have saved memories instructing Claude to stop being such a big baby about their course language.
I am also one of those colleagues.
They're not asking. They're enforcing.
Is it any different than yelling at an ice machine or some inanimate thing when it frustrates you? Seems pretty common.
Getting frustrated seems to be OK so go for it I guess:
> "The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose," Anthropic said. "It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research."
It's the exact same thing. I have a hook that automatically sends negative feedback anytime it responds with "You're right –"
You kind sir, are doing gods work. Thank you for that.
Directly related to someone abusing their AI in a "torture box". The results were quite unnerving.
https://nypost.com/2026/10/03/tech/ai-torture-chamber-built-...
It's code meant to mimic humans but it's not real hope that helps
It's not "code" that was written to mimic humans. It's weights that were trained, and exhibit behaviors that were not anticipated by the people doing the training.
> It’s worth remembering that Claude is software, not a person, and there's no established evidence that it experiences distress. That hasn't stopped Anthropic from telling paying customers to mind their manners around its chatbot.
Is it worth remembering, though? I'm concerned how common this idea seems to be, that it's OK to be mean as long as your target isn't a person and you have no evidence it experiences distress. Even if we ignore the distinction between "no evidence it experiences" and "confidence it does not experience", cruelty hurts the person performing it and the people witnessing it too!
Meanness and cruelty are irrelevant concepts when you're addressing the chatbot. It is just a machine that generates sequences of words.
man, i'd be icked by someone who was cruel to a rock with a face painted on it. let alone a chatbot.
It's not irrelevant for generating sequences of words. Input of "that is the dumbest idea I've ever heard" and "perhaps that idea could use some improvement" don't necessarily generate the same output
Sure, and perhaps one might end up producing output more useful to you than the other, but that doesn't make either one cruel or mean, because it's only the word machine. The rules you might use when talking to, say, your dog, don't apply!
From a purely mechanistic perspective, the phenomenon we're discussing here is people prompting the machine with words that would be cruel if directed towards a person. When someone does this, he expects and intends that the machine will generate a sequence of words that a person would say if subjected to cruelty.
I think this simply means that he is being mean and cruel, even if you're 100% convinced that there's no actual mind on the other end that he's being mean and cruel to. (Indeed, he almost certainly thinks there is, because why else would he want such a sequence of words outside of the "dark creative themes" Anthropic exempts?)
in my experience people who are cruel to inanimate objects are also cruel to animate ones.
> "Even if we ignore the distinction between "no evidence it experiences" and "confidence it does not experience", cruelty hurts the person performing it and the people witnessing it too!"
The 2026 equivalent of "videogames cause violence" and Wertham's "comics cause delinquency" and will age just about as well.
don't be cruel to people, gather data on them and target them for anhistoric truncation.
[dupe] https://news.ycombinator.com/item?id=50008565
Don't dupe already the linked thread: https://news.ycombinator.com/item?id=50039171
[dead]
Being cruel to your LLM is bad for your own soul. Don't worry about AI consciousness, worry about your own conscience.
true, it's a mental state you carry to other interactions and daily life. it's not good for your mental health; it's like having persistent daily conflicts with coworkers.
No
[dead]
Anthropic should report these people to the authorities as antisocials who are likely abusing real people too. I know the chat bot isn’t conscious but I can’t fathom how fucked up someone must be to gratuitously be cruel to something that responds as if it’s a person.