On November 12, Anthropic’s usage policy gains a rule against “sustained and needless abusive or cruel behavior toward our models.” It makes an easy headline, and almost none of it matters to a business buyer. Nobody on a 40-person team is going to be removed from Claude for snapping at a chatbot.
The rest of the update is what needs your attention. The reported changes tighten what staff can do with Claude in health, legal, financial and livelihood decisions, and they land on the same date.
Key points
- What does the Anthropic usage policy now say about cruelty? It says users may not repeatedly subject Claude to sustained and needless cruelty.
- Which changes reach a small firm’s own use of Claude? The ones on high-risk uses. According to RuntimeWire’s summary, for health, legal, financial, livelihood or essential-service decisions, a qualified human must be able to review and change Claude’s recommendations, and the people affected must be told that AI was used.
- Who in a small firm is on the hook? The account holder, in practice. An operator that accepts the policy accepts responsibility for how its staff use the tool, which is why a policy change is a staffing question.
- What should you do before November 12? Five weeks is enough. Do these in order.
What does the Anthropic usage policy now say about cruelty?
It says users may not repeatedly subject Claude to sustained and needless cruelty. The reporting on the update says frustration, pushback, dark fiction, testing and research are still allowed, and that ending the conversation stays the main response. Anthropic has not published examples of what counts, so the edge is unclear.
The line is in the policy text itself, with November 12 as the effective date. For a business, the practical risk is a chat that ends, not an account that disappears. We have not seen any statement of automatic account bans.
Which changes reach a small firm’s own use of Claude?
The ones on high-risk uses. According to RuntimeWire’s summary, for health, legal, financial, livelihood or essential-service decisions, a qualified human must be able to review and change Claude’s recommendations, and the people affected must be told that AI was used. That reads as a rule about the output, not the tool.
Anthropic’s August 2025 update applied a similar requirement only to consumer-facing outputs and excluded business-to-business use. Whether the new wording keeps that carve-out is not clear from the coverage we found, and we would read the final text before assuming it does. If a firm uses Claude to draft loan decisions, screen job applicants or summarize a client’s legal exposure, this is the paragraph to read. Open models have their own rules, and the Muse model licence follows every copy you make.
Who in a small firm is on the hook?
The account holder, in practice. An operator that accepts the policy accepts responsibility for how its staff use the tool, which is why a policy change is a staffing question. A hiring screen built in a spreadsheet by one enthusiastic coordinator is still your use of Claude.
We made a similar point about the Claude watermark: a vendor rule only helps you if you write your own rule next to it, and our look at agent controls found few firms have written theirs. Our review of the FTC probe and vendor contracts said much the same about paperwork. The vendor sets a floor, and your firm still has to write what happens above it.
What should you do before November 12?
Five weeks is enough if you work in order. List every place staff use Claude for a decision about a person, name the human who reviews each output, decide how affected people are told, and read the final policy text when it lands. None of it needs outside help. Do these in order:
Checklist
- List every place staff use Claude for a decision about a person, a loan, a diagnosis or a legal position. Ask them. The list is usually longer than the owner thinks.
- For each one, name the human who reviews the output and can change it. Write the name down.
- Decide how affected people are told. A line on the form or in the email is usually enough.
- Read the final policy text on or before the date, since the summaries differ at the edges.
Two further changes in the reported update are worth a glance. One covers connected hardware, where operators must be able to observe and stop equipment capable of causing injury. The other limits surveillance and criminal-justice uses. Neither applies to a typical office, and a warehouse running robots or cameras should read them.
The sceptic’s view is that a usage policy is a document, and the vendor enforces it rarely. That is probably right today. It is also the document a vendor, an insurer or a regulator will point to after something goes wrong, and “we did not read it” is a poor answer in all three cases.
Anthropic has not said how it will enforce the high-risk rules against business customers. Start with the list in step one, this week, and the rest follows from it.
Frequently asked questions
When does the Anthropic usage policy change?
The cruelty rule and the other reported changes take effect on November 12, 2026, according to Anthropic’s policy page. Anthropic announced the update on October 8. The money behind that policy is in Anthropic’s IPO filing, where the loss is mostly paper.
Does the cruelty rule affect businesses?
Almost never. It targets sustained, needless cruelty toward Claude, and the reporting says frustration, testing and dark fiction are still allowed. Ending the chat is the main response.
What should a small business do about the high-risk use rules?
List where staff use Claude for decisions about people, money, health or legal matters, name a human reviewer for each, and decide how affected people are told that AI was used. Provenance is the other open question, and the Claude watermark won’t settle who wrote what.
Last fact-checked October 9, 2026. Sources are linked on the claims they support.
