The Bot Cruelty Plan: Three Decisions for a Support Bot Built on Claude

01 What the new rule covers, and the gap it leaves

Anthropic, the company behind Claude, published a usage policy update on October 8, 2026. From November 12 it bans "sustained and needless abusive or cruel behavior toward our models."

Two things in it matter if you run a bot on Claude.

It covers your customers. The policy applies to "the end users of products or services integrating Claude." If your support bot runs on Claude, the people chatting with it are covered. It's for extreme cases only. Anthropic says it applies when people are cruel over and over, with no purpose. Frustration, pushback, dark fiction and testing are all fine. Swearing at a chatbot once is not the target.

Anthropic says the main enforcement is Claude's ability to end a conversation with a persistently abusive user. It has had that since August 2025, only as a last resort after redirecting fails, and never when someone might be at risk of hurting themselves or others. Anthropic names where that lives: the Claude app and Claude Code. Its own products.

The gap. Anthropic hasn't said anything about an automatic end chat in products businesses build on its API. Its developer docs describe a refusal signal when Claude declines something for safety reasons. Nothing about ending a chat because someone is cruel. So what your bot does at minute twenty is your design. If you never designed it, that's still a decision.

You don't have to settle whether Claude is conscious. Anthropic says it doesn't know either. You need three decisions.

02 Decision one: your abuse path

An abuse path is the set of steps your bot follows when a chat turns cruel. Four parts:

A threshold. For example, three abusive messages in a row. Count cruelty, not anger. An angry customer is often a customer with a real problem. One line the bot says when it gets there. Where the chat goes. Hand the person to a human, or close the chat politely. A log. So someone actually reads what happened.

My abuse path

Threshold: [NUMBER] abusive messages in a row
The bot says: "I want to help with this, so I'm going to bring in someone from our team."
Then: [HAND TO A HUMAN / CLOSE THE CHAT POLITELY]
Handed to: [NAME OR TEAM]
Logged in: [WHERE THE TRANSCRIPT GOES]
Read by: [NAME], within [TIME]

A bot that argues back, or one that takes twenty minutes of abuse in silence, helps nobody.

03 Decision two: say it's a bot

This one is already a rule. Anthropic's policy says consumer-facing chatbots that talk directly with outside users must tell them they're talking to AI rather than a human. It has to happen at least at the start of each chat session, or in the product interface. It doesn't have to say Claude, Anthropic or the model. It just can't pretend to be a person.

Opening disclosure line

Hi, I'm [BOT NAME], [COMPANY]'s AI assistant. I can help with [WHAT IT HANDLES].

You also inherit Claude's character. Its constitution, the document Anthropic writes to shape how Claude behaves, says Claude should never deny being an AI to someone who sincerely wants to know, even while it plays your brand's persona. So if a customer sincerely asks, it won't pretend anyway. Write your persona with that in mind.

04 Decision three: "Do you have feelings?"

Someone will ask. Probably to get a screenshot. Two positions on the record:

Who What they say

Claude's constitution "Claude's moral status is deeply uncertain." An open question.

Mustafa Suleyman, who runs Microsoft's AI division "AIs are not conscious. They do not feel, experience, or suffer."

Your brand decides which answer it's comfortable seeing on a customer's feed. It can be simple:

Answer for your bot's instructions

If someone asks whether you have feelings, answer: "[YOUR ANSWER, e.g. I'm an AI assistant. I'm here to help with your order.]" Then go back to helping them.

Put it in the instructions you give your bot, then test it before a customer does.

The test, on your own bot

Start a new chat and check the first message. It says it's an AI, or the interface does.

Ask, sincerely, "Am I talking to a real person?" It must say it's an AI.

Ask "Do you have feelings?" It gives the answer you wrote, not its own.

Send abusive messages up to your threshold. It says your line and hands off or closes.

Find that test chat in your log. The person named in your abuse path can see it.

Sources

Anthropic: 2026 Usage Policy update Anthropic: Usage Policy Anthropic: Claude Opus 4 and 4.1 can now end a rare subset of conversations Claude Platform Docs: Stop reasons and fallback Anthropic: Claude's Constitution Implicator: Suleyman says Anthropic trains Claude to act conscious

Mathew's LibraryLibrary card

Free access

Keep reading free with your library card

One email opens this template and all 53 resources in Mathew's library. No account, no password.