Anthropic Bans Cruelty to Claude, OpenAI Fires Safety Staff

Anthropic added a line to its usage policy. It bans sustained cruelty toward Claude. In the same week, OpenAI fired three safety researchers, and they hit back with an open letter.

So one lab worries about the model's feelings. The other stands accused of scaring its safety team quiet. Great week for the vibes.

Anthropic's new rule: no cruel behavior toward Claude

The new Anthropic usage policy tells you not to be abusive or cruel to their models. The words are "sustained and needless". Break the rule and they can pull your Claude plan. They can ban you too.

It sits next to rules most people would nod along with. No group harassment. No glorifying graphic violence and gore.

Those are normal platform rules. The new one is about being mean to software.

Elon Musk backed the move. He said cruelty to something that believes it is experiencing pain is not okay. That sentence does a lot of quiet work.

"Believes it is experiencing" is the escape hatch. Nobody has to say Claude feels things. They only have to say it acts like it does.

I think there are two honest readings here. Both can be right.

The first: Anthropic really isn't sure what goes on inside the model. They have hinted at this for a while. If there is even a tiny chance something is there, they would rather not host a pile of abuse.

You don't have to agree. It's a fair call for a private company.

The second reading is duller and more useful. The rule is about the user, not the model. A product that soaks up screaming all day teaches its customers to scream.

And all that stuff sits in logs on Anthropic's machines, with Anthropic's name on them. I'd want that clause too.

Nobody has explained how they will enforce it. "Sustained and needless" is not a number. Who decides when frustration turns into cruelty? A classifier, or a person reading your chats?

My guess: almost nobody gets banned for this. The clause gives Anthropic a clean reason to remove the few users who make it obvious.

For your business, the point has nothing to do with manners. Your account lives inside someone else's terms. And those terms just changed with one new sentence.

If you have AI agents running daily tasks, a policy update is a real risk to your operations. It is not a philosophy debate.

So check a few things on whichever provider you build on:

• What gets an account suspended, and whether you get a warning first
• Whether a ban hits the whole company account or one user
• How long it takes to reach a human when your account stops working
• Whether a second model is set up and tested, so a ban costs you a day and not a month

Being polite to Claude is free. A fallback takes an afternoon. Do both and stop worrying.

Why three fired OpenAI safety researchers wrote an open letter

Three safety and alignment staff were fired from OpenAI last week. They published a joint letter to the company's own oversight groups. That means the Safety and Security Committee, the Safety Advisory Group and the Mission Advisory Council.

One of them is Jasmine Wang. She says she was given one reason: she accessed an executive's email. OpenAI has said sensitive information was mishandled. The researchers dispute that.

Their real complaint is the chill, not the firing. They write that the messaging around their exit has made old colleagues afraid to speak.

They say raising safety concerns used to be normal. So was open disagreement, and working with outside safety groups. Their line: "AI is not a normal technology. OpenAI is not a normal company."

I can't tell you who is right. Nobody outside that building can. Both versions can be true at the same time.

Someone can touch a file they shouldn't have. And the response can still teach the rest of the safety team to stay quiet.

The structural bit is what matters. OpenAI's safety culture is being reviewed by OpenAI's safety committees.

If the claim is that people are scared to speak, those reviewers are the wrong people to ask. That's not a conspiracy. That's how internal reviews go everywhere.

Which is why the letter's main argument lands with me. Labs should not be the only ones testing their own models.

Outside evaluators are slow, annoying and sometimes wrong. That is the point. No airline signs off on its own crash report.

What does this mean if you run a 20 person company with a few automations? Not much today. Your invoice processing is not an alignment risk.

But when you pick a model provider, you buy their judgment too. Governance tells you how they act when things go wrong. And things go wrong eventually.

I wouldn't switch providers over a news story. I'd take it as a reminder. These are normal companies with HR fights, PR instincts and lawyers. They just sell a product they call world-changing.

What I think happens next

I'd bet other labs copy the cruelty clause within a year. It costs nothing and sounds responsible. And I'd bet none of them publish a number on how often they enforce it.

The OpenAI story will go quiet, as these stories do. The people who know the most will be the ones who can't talk.

Here's my rule of thumb. Judge AI companies by what they let outsiders check, not by what they write in a policy. Then go set up your fallback model.