Being mean to AI could get you banned as Anthropic introduces new rules against abusing Claude
WTF?! You might have heard the one about being nice to chatbots in case they remember you once the machines take over. It seems Anthropic is heeding this warning, barring users from being excessively abusive or cruel toward its AI models.
Anthropic’s first changes to its usage policy in over a year include new restrictions on abusive behavior. The updated policy prohibits “sustained and needless abusive or cruel behavior” toward its models.
Don’t worry if you’re a Claude user who occasionally gets annoyed at the chatbot and throws a few expletives its way. The policy update only applies in extreme cases, in which users repeatedly act cruelly toward the models for no discernible reason.
I don’t and wouldn’t bother, but I don’t see the point. If this is a sign of sociopathy, removing the account doesn’t cure the mental illness. If it’s being fed back into training data, stop. This seems needless, though it’s likely to lead to increasingly strict but completely vague usage requirements. How do I know exactly what behavior falls under this? So I have to over-moderate just in case.
Not in this case but this could lead to that sort of thing.
If they can pick out that kind of behavior they can filter it for training. This is probably more to protect them legally so when they shut down a user they have something to point to.
That people are getting angry and insulting to an LLM shows they don't understand the technology (true of most of technology really), and probably think it's actually understanding them or even caring. Really embarrassing and sad when you picture an LLM output like a type of a mirror, GIGO. And the mental health aspect of users is a huge issue that's not being talked about enough, and also being exploited to sell a solution.
I curse at inanimate objects all the time. Cursing at an LLM is like cursing at the coffee table when I stub my toe except the coffee table doesn’t respond.
That would be the easy solution. Number one, don't use such responses for training, but I think they aren't. Two, just reply with things thrown at it with "..." And three, the really angry people end up not using it, which benefits everyone, even the angry person.
The problem is they’re not managing the training data. There is no one sitting there making sure all the data that gets retained for training is “good” data, they’re just expecting everything it sees to be “good” data.
Just like the data hoovered up from the internet that was used to build the models wasn’t vetted either, which is why AI goes wrong SO often…
It’s fairly simple to just run automated analysis on data quality once you already have a decent LLM though. Claude would definitely be able to.
And they have people doing that manually too, or perhaps it was OpenAI. Someone got angry at their underpaid contractors for using AI to create training data lol
I think it might be more the individual mirroring they tend to do after a while. If you are consistently mean to them, they eventually mirror it back to you, and then people get funny screenshots of them saying mean things and put it on social media.
Or even just in general starting a downward spiral feedback loop where you both keep being slowly meaner and meaner to eachother until it’s starting to actually hurt you long term.
I wish the AIs would be abusive. That would be so funny if one day there’s an update and they start telling all these people addicted to the chatbots like “Go outside. Please.” Or someone asks it a stupid question and it’s just like, “You’ve gotta be fucking with me, right?”
I’m gonna say it, I think this is a good rule for anything that mimics human interaction like video games do. Not for the game but for the people using it to not get used to just running over whatever is in their way.
-Shamelessly stolen from top level comment on this thread
I have run over thousands of pedestrians in GTA3-5. I have beat hundreds of people to death with a bat. I have gone on hundreds of shooting sprees where I sniped drivers in cars, pedestrians, launched rockets at them, thrown grenades at them.
Not once did I have the urge to act violently to a human. Not once did I have the urge to shoot the moon so it gets bigger either.
The only FPS video games I find objectionable are ones where you play as the good guy in American army camo and shoot at “terrorists” dressed in shmaghs and the setting happens to be their country.
My claude ignores half the stuff I define in it’s claud.md once it goes over ~200 lines, and I need it to know more important things like “don’t try to locally verify with chromium” so it doesn’t spend 20 minutes + $10 in tokens rediscovering I have removed it’s permissions to accomplish that.
People are already anthropomorphizing AI. That’s a good chunk of the problem. If you get used to being shitty to an AI that takes it, then you’re a lot more likely to be shitty when you go touch grass.
This is also the best argument I ever heard against AI friend/relationship apps. It’s teaching you a certain style of communication which may well prove catastrophic in real life relationships.
Given the many people that send death threats to actors for the actions of the characters they portray… you might be underestimating the number of “stable” people on this planet.
Sadly, I think you might be right considering how fast semantic pareidolia is becoming the norm. It’s only a matter of time before the war crimes I’ve committed in Medieval 2 Total War are reified and I’m put on trial.
That’s just a numbers game. Even if only 0.01% people are that crazy, that’s still over 30k people in just the USA. And when you’re only hearing about the worst of the worst, it can easily throw off your perception of the overall state of the population as a whole.
Honestly, this is a counter-example to what you’re saying. Despite how many people have access to the internet, world renowned celebrities still only get so much hate-mail. Most people don’t do this shit, but obviously a small percentage is still a large population when multiplied by the entire global population.
I’m not sure what abuse means in this situation. Maybe abuse in overburdening the system, but words do not cause harm as we know it to the algorithms. If we assume that these algorithms can understand anything like we do, there is no reason to think talking they have the same criteria for abuse that we do. For all we know belittling and yelling at them is welcomed and taking calmly to them is actually abuse. Though I think it has no concept of abuse to even start with. It is just a stochastic parrot that says and responds in a way based on whatever is in the training set.
This happens whenever you’re talking about any topic that the AI model provider has decided arbitrarily they don’t want to do.
I got temporarily banned off deep-seek the other day because I was curious about how well generation technology has gotten for self-hosting. So I asked it about setting up Comfy UI because I had never used it before. Woke up the next morning to a your account has been suspended because we have deemed you broke terms of service. In order to appeal you need a Chinese phone number, it made me chuckle.
Considering the only two questions I asked it was about comfy UI or how to make a recipe similar to the now discontinued flaming hot cheetos Mountain Dew(I missed the production window sadly!)
I’m leaning towards the comfy UI being the reason.
Yeah basically the whole point of capitalism is that private property has more rights than literally anything else. It is treated as sacred. Ironically the dehumanizing thing is usually said to be when they treat women like property but if they actually treated women like property then women would probably be safer.
Kinda madly irritating that this excludes the whole charming population of Scotland to hone their great cursing skills because these motherfucking clankers which have as much of a soul as a thermonuclear bomb detonator, as much sincerity as a Russian landmine camouflaged as a child’s toy, as much empathy as a cruise missile hitting a girls school, and as much of a hearth as the boss of Oracle firing 10,000 people with a pencil stroke, and moreovee are too dumb to screw a lightbulb socket are allegedly too sensitive and dainty to receive what people are actually thinking…
It has nothing to do with feelings. Talking in a very hostile way is a strategy to get the model to go around it’s set rules, and that’s what they are trying to prevent. Yes it’s ridiculous, but it had nothing to do with caring about the model.
It looks like bad training data, as they harvest all of their subscribers.
It looks like 2013 Microsoft having to unplug their early AI project from Twitter within 16 hours as the general public very deliberately turned it into a Neo-Nazi.
It looks like conserving tokens to conserve energy costs not expended on dissidents.
Makes sense then, rather than improve the system to not just cave to rude words they tell people they can’t use rude words… Just more proof it’s all a big joke.
Makes sense. Get it into an adversarial kind of context and it will predict more text that goes against other rules “established” in the context.
There will be other paths to do this. Like a context where there’s a lot of suspicion for other parts of the context would be one of my top guesses. Given the way they work, you could gaslight the shit out of them, since they aren’t an entity that has any memory of its actions. You can even edit the context to modify its responses, which will affect the tokens it predicts going forward.
Though I’m curious how this will be handled by companies that expose access to claude to anonymous people on their website, like ddg.
That’s just a part of the process, once they are liquidated we then atomize them into a gas then we compress them into liquid, we then use that to fuel cargo ships.
