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.
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 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!
> "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 will age just about as well.
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?)
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.
They are asking this because, at scale, this behavior probably has some negative effect on the post-training process.
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.
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.
http://nicehole.urbanup.com/8485154
Frontier AI folks may be crazy, but not crazy enough to believe people read usage policies :-)
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."
And they can't disclose it, since then they can be found responsible for such negative influence and be liable for the damages.
• 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
I am also one of those colleagues.
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)
> "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."
https://nypost.com/2026/10/03/tech/ai-torture-chamber-built-...
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!
The 2026 equivalent of "videogames cause violence" and will age just about as well.
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?)