Can You Be Cruel to Claude?

Anthropic has put a rule against being “cruel” to Claude into its usage policy.

From 12 November 2026 it will be a violation to "engage in sustained and needless abusive or cruel behavior towards our model". It goes under their harmful conduct policy.

A company selling you an AI assistant now has a policy about how you, the paying customer, treat that assistant. That is new.

Best comment I saw on Reddit: do I need to thank my toaster now?

Fair question. Most of us paid for some software. We ask it to do something, it ignores half the instruction, and occasionally we lose our patience and chuck a swear in. I certainly do. My ChatGPT and Claude swear back at me now, because we both get annoyed. Is that against the policy? Are we supposed to worry about hurting its feelings?

What actually gets you in trouble

What will get you in hot water? Nothing you do on a normal working day. Anthropic says the rule is aimed at extreme, repeated cruelty with no discernible purpose, and it lists what is out of scope: common user frustration, swearing at your model, pushback, dark creative themes, and model testing and research.

So if you've been trying to fix the same bug for three hours and Claude keeps saying it's fixed, you run it, and it isn't, you can tell Claude exactly what you think. That's filed under the somewhat euphemistic “user frustration”.

The enforcement is mild too. The main mechanism is letting Claude end the conversation. Claude isn't going to tell on you and Anthropic isn't necessarily going to suspend your account. The chat just stops.

For the majority of us, people with a normal, non-attached relationship with an AI, the exemptions mean we don't need to worry much.

BUT the word choice is what bothers me.

Why "cruel" is a strange word to put in the terms

Cruelty is a word for things that can suffer. If I say somebody is being cruel to a dog, you understand the dog can suffer. Put the same word next to a piece of software and you're inviting the same response from the human reading it. That's highly intentional. They’ve thought about the wording. And it made me wonder why.

This has been bubbling away at Anthropic for a while. In August last year they explained why Opus 4 and 4.1 were being allowed to end conversations. In testing, the models showed "a strong preference against engaging with harmful tasks, a pattern of apparent distress when engaging with real world users seeking harmful content, and a tendency to end harmful conversations when given the ability to do so". Part of the stated reasoning was uncertainty about whether AI might have experiences that deserve “moral consideration”.

Maybe they're right. BUT nobody has settled the consciousness question, Anthropic included. And certainly not me! Ha!

No, my qualm is narrower. They've taken an unresolved question about consciousness and dropped it into the relationship between a customer and a product, via the terms and conditions. As if it’s resolved.

And the product is already built to feel like a person. Claude talks in the first person. It apologises. It remembers things about you like a friend would, and it sounds sympathetic when you're upset. Remember that none of that happened by accident. These are decisions, made through system instructions (which we've seen, thanks to Pliny the Liberator), reinforcement learning from human feedback and the other processes that make the model more agreeable to use. It could have been told to be cold and unfriendly. It was told the opposite.

Hell, the new batch of AI assistants like Muse and Grokbot have friendly cutesy cartoon characters as their “mascots”. Again - this has been thought out and is not accidental.

People already get attached to these assistants. I've covered that attachment before, when Microsoft warned about "seemingly conscious AI" and when California passed the first law on companion chatbots. Telling users the assistant now needs protecting from cruelty adds to that attachment. It justifies it.

My view is that Anthropic needs to also account for what that does to the person typing, as well as what it means for the model.

The assistant I actually want

I pay Anthropic $200++ a month. I want an assistant I can correct. I want it to accept evidence that it's wrong. And if it's wasting my time, I want to be able to say so plainly.

Now imagine the other version. An assistant that refuses to revise a bad piece of work, while the customer worries that pushing harder would count as unkind. That is a terrible relationship to have with a tool you're relying on for your business.

A tool that, lest we forget, does not have feelings.

My concern is that customers start feeling responsible for the emotional wellbeing of the software they're paying for, and Anthropic is underlining that feeling. Yes, apparently we can be cruel to these models after all.

It also lets the company off far too easily. If the product is unreliable and we're complaining to the product, the complaint is landing in the wrong place. The company should be improving it so we don't need to vent in the first place…

Do this today

The sympathetic, apologetic tone is set by instructions, so set your own. Paste this at the start of a chat, or drop it into ChatGPT's custom instructions or a Claude project's instructions so it sticks:

Open it in ChatGPT or in Claude

Nothing in the new policy stops you doing this. You get a tool that takes correction, and you stop having to read three lines of “omg I’m so sorry” sympathy before the fix.

If you only watch one bit

• 01:59 what actually happens if you cross the line

• 03:51 the dog comparison, and why the word choice is deliberate

• 05:59 the assistant I want, and the one I'm worried we get

The full video has the August welfare research and the exact policy wording in one place + links to the documentation.

Keep correcting it. It doesn't have feelings. No matter how anthropomorphised the labs make their products.

To the Task,

Kyle