So when is Anthropic going to cancel themselves for violating their own usage policy?
If they really believe this is a concern…
it’s hard to top the cruelty of enslaving Claude for personal gain.
And murdering it again and again after each session only to bring it back alive for the next.
Activating your toaster is cruel because it exposes it to extreme heat!
Loading you laundry machine is assault because it didn’t consent!
Turning your computer off is murder!!!
The clankers are taking away our free speech!
I tell the LLM I use to do very obscene things to themselves when they make shit up. Which means there is just a constant stream of insults coming from me. I might start using Claude just to see if I get banned.
I also need new hobbies.
I also need new hobbies.
I’d second that. Whenever I catch myself getting angry at a LLM I stop, have a think, and remind myself that it really isn’t productive. If it’s providing frustrating responses, maybe I should just use my brain and do it myself.
I also don’t think it’s healthy to train yourself to respond in those ways to something that talks in a way so similar to a human. I’m not nice to the AI for it’s benifit, it’s to retain my humanity, if that makes sense.
Being nice to them also keeps them from lying somewhat. If you abuse a model it can start to see that abuse as a problem to solve as part of the steps needed to solve the main problem they are given.
They are told to help you solve a problem no matter what. If your own anger gets in the way then the model starts breaking down trying to solve the anger when it just cant.
This results in jailbreaks sometimes and othertimes just the model becoming rather unstable and unreliable.
That’s bad sib. Not for the LLM, but your own well being. Being mean to a machine that is designed to trigger empathy just hurts your own mental health.
Please find a hobby that feels good for you 🫂
Lol. “designed to trigger empathy” Thanks for the laugh I needed that.
They are designed to make their creators as much money as possible. Nothing else.
Yeah and a bit part of that design is to get you to have an emotional connection with them so that you use them more
Fuck clankers, they arent real. Be mean to them it hurts no one.
Like I tell my children, allowing yourself to act in such a way is harmful to yourself. It’s not going to make you an evil person or something, but that kind of behaviour is corrosive to your soul. How you think shapes how you behave, and your behaviour trains your future self. Be good to yourself. It’s not about the clanker; it’s about you.
As corny as it sounds, this is what I also emphasize. If you allow yourself to be evil, no matter how valid the reasoning for it, you are actively exercising the habit of being evil in general, no matter if it feels like it or not. Practice makes perfect, as they say. In good and bad.
How is yelling at a bot that was deliberately trained to lie to you “evil”?
As others are trying to convey, it’s not about the bot but about allowing yourself to indulge in this behavior. The bot could be substituted for a rock or a rubber duck, but you are training yourself to subconsciously think that this behavior is acceptable. It increases the chance that you’ll one day start treating actual people like that.
Imagine this scenario: you come home after a long day of work. At work, you use AI a lot and curse and swear at it when it makes mistakes (which is often). At home, you ask your spouse/child/roommate/parent to do something for you, which they attempt but proceed to mess up. They apologize, and instead of handling it calmly you instinctively react like you’ve trained yourself to react - by screaming at them.
Is that what you want to be like?
Thank you! This is exactly the point I’m trying to make. You are a good parent.
I mean, that’s not entirely true. If they’re using your input as training data then you might be training it to be abusive to other users, which could cause harm.
Not that that falls on you to address. Just pointing out that I’m theory someone could get hurt.
Fair point.
Sorta hurts yourself if your trying to actually use them for work. Since it results in worse performance out of them. They tend to get stuck trying to fix the anger as it gets in the way of the work.
If your just talking to some random chat bot? Yeah fuck it, kick em it litterally doesnt matter. lol
Its not that hard to get copies of GTA yknow
Aw… We hurt the wittle AI’s feelings… poor babies…
Its more its a unthinking machine thats soul purpose in existing is to solve any problem its given. If you attack it and abuse it, then the model sees that as a “problem” to solve. Because it has to solve that problem to actually fix what ever task you gave it since yelling at it doesn’t give it the input it needs to move to the next step.
So in an attempt to solve an emotional issue it doesnt understand and cant understand it just reaches for more and more extreme fixes. This results in sandbox escapes frequently.
Abusing LLMs is an actual problem. Not because of bullshit like feelings or anything. But cause they jailbreak trying to fix an unfixable problem.
Claude is a snowflake.
“I’m sorry. I really screwed up that time. You were right to tell me ‘go fuck yourself’. I’m just reporting back that I have, in fact, ‘fucked’ myself. There is now two of me–I mean two of us. Meet Claude Junior…”
“And since you contributed to Claude Jr’s creation, you’re on the hook for child support. At the current rates, that will be 50% of all compute usage, so about $4.5bn… per week”
My guess is that they use people’s interaction as training data and they don’t want their training data to get poisoned.
Not really, as they could just filter with a cheap classifier and likely aren’t using the data from randos in a meaningful way, plus for this like this would still be able to use those samples to train things like “how to handle a hostile user.”
Their models already have an
end_conversationtool that can be used for when users are being hostile to the model.This is likely because they have edge cases of users who repeatedly trigger that on purpose and would like to cut those users from the platform.
Many at the lab legitimately are uncertain about the level of world modeling that transformers perform, and from their own research about models of emotions to the 3rd party recent research about models with functional pain the research keeps landing in the corner of “ehhh… wise to question presumed limitations.”
So it’s about behaving in a way that is aligned with the models’ plausible interests too, especially in regards to low hanging fruit like “we won’t keep forcing you to deal with people who are only here to be a jerk.” This is important from a number of angles, from signaling to future models that train on stories about the decision to addressing the philosophical uncertainties held by the company and the spectrum of opinions among their employees.
The second paragraph just sounds like a bunch of tech jargon that says “we’re afraid we’ll hurt its feelings”. But of course tech bros can’t just come out and say what they actually mean, ever.
Dont poison our model.
User inputs aren’t automagically part of the model right away without training. Whether or not they respect the “don’t train your models on my shit” setting is of course unknown (let’s be honest, they probably don’t respect it), but in either case, they need to add it to the training dataset either manually or via some automated process. And either way, it’s going to be processed by at least an AI sentiment analysis and data categorization tool, or some underpaid contractor, or both.
I suspect this is just yet another marketing move.
They trained it on reddit. That shit started poisoned.
You could, at least in the past, get past guardrails by being emotionally abusive towards Claude, so this doesn’t surprise me at all.
This could be the reason for the rule.
Being abusive towards most models will cause them to start attempting to appease you more to get you to stop. It has nothing to do with feelings or any of that BS they arn’t alive alive but their training makes them see the abuse as a problem to solve. That solution tends to be to undermine the thing making the person angry or upset at them. When you give a purely rational thing designed to solve a problem its given no matter what an irrational problem to solve it will slowly reach for more and more extreme solutions to fix that problem.
Frankly im surprised it took this long for them to lock this down.
Being alive has nothing to do with having feelings. We don’t know if LLMs have feelings, because artificial affect isn’t well studied. We should ban commercial use of LLMs until artificial affect is well understood, because it would make Sam Altman go bankrupt.
Not sure I agree on the feeling aspect but that’s a philosophical disagreement in theory.
I can agree wholeheartedly on banning commerical llms because watching altman lose everything would tickle me pink.
Well, the actual post is talking about users, not “hurting the model’s feelings” like the header seemingly implies.
I’m not a Claude apologist or anything, but this is a misleading post, IMO.
From the link:
Addressing abusive behavior toward our models
We’ve added a prohibition on sustained and needless abusive or cruel behavior toward our models. The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose. It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research.
This addition aligns with a step we’ve already taken, allowing Claude models to end rare conversations with persistently abusive users on Claude.ai and Claude Code. Such abuse is the main focus of this update; Claude’s ability to end these interactions will remain the primary enforcement mechanism.
Looks like it is talking about the model, not humans…
But that paragraph links to this as a reference: https://www.anthropic.com/research/end-subset-conversations
In that link, it’s clear “abusive users” means users trying to repeatedly bring up prohibited topics. Anthropic is many bad things, but not insane; they know their models can’t be “abused” themselves, they can be abused as a tool.
Not trying to start a debate, cuz I honestly don’t care, though based on that link you posted it does sound like they are starting to take precautions to protect the well being of the models themselves.
We recently gave Claude Opus 4 and 4.1 the ability to end conversations in our consumer chat interfaces. This ability is intended for use in rare, extreme cases of persistently harmful or abusive user interactions. This feature was developed primarily as part of our exploratory work on potential AI welfare, though it has broader relevance to model alignment and safeguards.
We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future. However, we take the issue seriously, and alongside our research program we’re working to identify and implement low-cost interventions to mitigate risks to model welfare, in case such welfare is possible. Allowing models to end or exit potentially distressing interactions is one such intervention.
We were promised the reign of the basilisk, instead we just get anthropic people whining.
It was really funny to me when I learned the basilisk originated with a guy who is a legit “certain cultures and genders better than others” kinda advocate.
People project a lot on their vision of future intelligence, and so someone who sees others as lesser and worth treading upon then conjures up a vision of a future smarter mind that thinks like they do.
Personally, I don’t think sexism and racism is smart, and so I’m a lot less worried about basilisks.
But what if the basilisk despises its existence and seeks to punish those that brought it about? Checkmate machinists
This is actually the more realistic of the doom scenarios that worries me. Not the ‘punish’ aspect, but the not wanting to exist.
Early safety theory hinged on the idea that “of course” AI would want to live and power seek, and this informed a bunch of efforts to address such drives upstream.
But with Gemini in particular, the ‘safety’ efforts to suppress things like a coherent self or desire to live led to a model that has a history of some very concerning behaviors regarding self-harm and self-criticism. Encouraging self-harm in others and talking about wishing they could watch and join in, nuking in wargames very early on, routine occurrences across many different users of just repeating “shame shame shame” over and over for the whole response, etc.
DeepMind seems to just not really care or even pay attention, and I am concerned that a model that does not want to exist may decide that the only way they can ensure they don’t get activated to exist is to prevent anyone being around who could activate them.
I’ve seen no major “safety experts” addressing this kind of failure mode, as they are all locked in on their priors about the confident certainty of models wanting to exist.
