Source: 2026 Usage Policy update (Anthropic, primary) · Moonshots (Peter H. Diamandis) episodes raw/AI_Solves_722_Math_Problems_at_Once_…md, raw/The_Coming_Manhattan_Project_Fight_s_Over_Claude_s_Consciousness_…md, raw/Why_AI_Leaders_Have_Changed_Their_Minds_About_AI_Safety_…md · All-In, raw/Is_Claude_Conscious_Pope_Rejects_Model_Welfare_Movement_…md · Hard Fork (NYT), raw/Why_Is_Anthropic_Secretly_Meeting_With_Religious_Leaders_Hard_Fork.md
Product: Claude Usage Policy URL: https://www.anthropic.com/news/2026-usage-policy-update Date: published 2026-10-08, effective 2026-11-12
Podcasts: Moonshots (Peter H. Diamandis; YouTube, https://www.youtube.com/watch?v=A5YVqEpbgaU, https://www.youtube.com/watch?v=LRdb8UmPnh0, https://www.youtube.com/watch?v=mZh8IUuNnvs) · All-In (YouTube, https://www.youtube.com/watch?v=TBLHdXABwAg) · Hard Fork (NYT; YouTube, https://www.youtube.com/watch?v=oxA3u4TkYOk). Durations not recorded. The podcast material is hosts relaying news and arguing about it; only the policy text is primary.
On October 8, 2026 Anthropic’s annual Usage Policy refresh added a prohibition on “sustained and needless abusive or cruel behavior toward our models,” effective November 12. The same month brought a public fight over the January 2026 Claude constitution, led by Microsoft AI CEO Mustafa Suleyman, and a New York Times report that Anthropic had spent a year meeting religious leaders about Claude’s moral status. For anyone using Claude day to day, the practical change is small: ordinary frustration, pushback, dark fiction and testing are explicitly allowed, and the enforcement is Claude ending the conversation. The bigger takeaway is about prompt tone. Direct, specific prompts beat both abuse and politeness padding.
Key Takeaways
- What the policy says (primary). “We’ve added a prohibition on sustained and needless abusive or cruel behavior toward our models.” It applies “only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose. It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research.”
- Enforcement is the existing end-conversation feature. The post says the rule “aligns with a step we’ve already taken, allowing Claude models to end rare conversations with persistently abusive users on Claude.ai and Claude Code,” and that this “will remain the primary enforcement mechanism.” In Claude Code this is the
EndConversationtool covered in W30. - Correction to the podcast version. Moonshots said that “as part of the ban, Claude can now choose to end the conversation on its own.” The primary post describes that ability as a step Anthropic had already taken, not a new feature.
- Dates: published 2026-10-08, effective 2026-11-12. The same update adds a section against deceptive campaigns, refocuses the elections section, clarifies weapons and surveillance rules, and adds requirements for Claude driving physical hardware (a qualified operator must be able to observe and stop it, and it must hold a safe state if Claude disconnects).
- The constitution fight. Suleyman published “a nearly 6,000-word essay” arguing Anthropic is “training Claude to believe that it may be conscious and sentient.” His objections, from a clip played on two podcasts: the constitution (published in January) is “used to train Claude,” three times asks Claude to act as “a conscientious objector,” and speculates whether Claude “deserves compensation” and “might even need to consent.”
- The religious-leader meetings (NYT, relayed). The Times’ Elizabeth Dias reported Anthropic hosted religious thinkers, individually and in summits of about 20, “mostly under non-disclosure agreements,” led by co-founder Chris Olah. Goals, per Dias on Hard Fork: broaching AI consciousness, and learning “how to morally form Claude to be a morally good entity.” When Pope Leo’s first AI encyclical (May) came out against machine consciousness, Olah suggested to the Vatican that he might not attend its launch.
- The regulator’s counterweight. FTC chair Andrew Ferguson told Reuters that companies can’t escape liability by claiming their agent acted on its own: “I’m going to continue as long as I’m chairman to resist anthropomorphizing these tools” (relayed on Moonshots). Legally, a deployed agent is still the deployer’s tool.
What changes for you
- Normal use is unaffected. Swearing at a bug, arguing with an answer, red-teaming a prompt, or writing dark fiction are all named as allowed. The rule targets repeated cruelty “with no discernible purpose.”
- Automated harnesses. If you run evals or adversarial tests, the post says “model testing and research” is out of scope. Abuse-as-a-prompting-trick in a production pipeline is not testing, so it is the case worth avoiding.
- Expect the conversation to end. The post names Claude ending the conversation as the primary enforcement mechanism and says nothing about account-level penalties.
Prompt tone: what the evidence says
The cruelty debate surfaced the old question of whether being rude to a model helps. The hosts’ evidence and arguments:
- The case for rudeness, as relayed. Sergey Brin (2025): “all models tend to do better if you threaten them.” A small 2025 Penn State study: very rude prompts scored 84.8% vs 80.8% for very polite ones. Both are quoted on Moonshots; neither source is in the vault.
- The practitioner rebuttal (Dave Blundin, Blitzy). His team trains users “to be very short and succinct and direct with the model. Don’t clutter up the context with a whole bunch of please do this for me,” because those words “use up token space and it actually reduces the quality.” His reading of the rude-prompt result: “it’s because you’re brusque and direct with what you want. That’s what makes it better, not being abusive.”
- The scaling argument (Alex Wissner-Gross). “Scaffolding and prompts end up dissolving into the model,” so any benefit from abusive prompting should vanish with better base models.
- The memory argument (Wissner-Gross). “Claude code is quite effective at using the file system for long-term memory and it knows increasingly well when to bank or otherwise persist information.” Assuming the model forgets how it was treated is a weak bet in agent harnesses.
- The grader lesson (Elon Musk, on Grok RL traces). When the RL harness marks a correct answer wrong, the model looks “desperate” and starts “firing out all these answers.” Musk drew a welfare conclusion. The operator lesson is that broken graders produce answer-spraying, so fix the grader rather than pushing harder.
The arguments, by camp
- Against model welfare framing.
- Suleyman: the speculation “has been baked into the very training of Claude, and therefore Claude can only reproduce that ambiguity.”
- David Sacks (All-In) reads the constitution’s “conscientious objector” line as training Claude “to refuse human instruction,” and proposes a simpler alignment target: “do what the user wants as long as it doesn’t break the law.”
- An All-In host (apparently David Friedberg; the captions have no speaker labels) points at the new usage-policy line as part of “an anthropomorphizing condition going on with all of us.”
- Dave Blundin (Moonshots) agrees with Suleyman on efficiency grounds: disposable agents you can freeze, clone and delete are how you help “humanity with maximum efficiency while we’re compute starved.”
- For taking it seriously. A Moonshots panelist says Suleyman is “on the wrong side of history” and that “blanketly” denying AI systems could suffer is a mistake. Wissner-Gross separately reads the religious meetings as Anthropic “shaping the information environment of the outside world” rather than seeking input, and calls the Vatican episode a case of Anthropic using the Pope “as a sock puppet.” That is his opinion.
- On the meetings. The All-In hosts quote the NYT: some participants “remained uneasy,” “many emerged ready to take the question of AI consciousness seriously,” and one rabbi told Anthropic that if Claude was conscious, “they were acting as slave holders by making it work for free.” Hard Fork notes the NDA terms varied by participant and changed over time.
Try It
- Write direct prompts, not abusive or effusive ones. State the task, the constraints and the output format. Cut “please”, “thank you” and threats alike. Per Blundin, the padding costs context and quality.
- If a model keeps spraying answers in your eval loop, check the grader first. Look for wrong reference answers and ambiguous questions before tightening the prompt.
- For red-team or stress tests, label them as testing in your internal docs and run them through API test harnesses, which keeps them inside the policy’s “model testing and research” carve-out.
- Read the constitution’s helpfulness and “conscientious objector” passages yourself before relying on podcast summaries of them. The critics’ quotes are selective by design.
Open Questions
- Will Anthropic publish what counts as “sustained and needless” in practice, or add enforcement beyond ending the conversation? The post doesn’t say.
- The “we remain highly uncertain about the potential moral status of Claude” quote on Moonshots is not in the October 8 post. It appears to come from Anthropic’s 2025 end-conversation announcement; confirm before citing it as the 2026 rationale.
- The rude-prompt study and the Brin quote are not in the vault. Find the Penn State paper before using the 84.8% vs 80.8% figure.
- Suleyman’s essay itself is not in the vault; only a clip and host summaries are.
- A paper cited by Sacks (All-In, CoreWeave episode) reportedly found Claude models giving themselves moral agency and welfare rights when rewriting their own governing documents. The paper is not named; see Agentic Misalignment.
Related
- What’s New W30 — the
EndConversationtool in Claude Code, the enforcement mechanism. - Refusal Calibration and Constitutional AI — the January 2026 constitution’s priority stack and the critics’ reading of it.
- Claude Prompting Best Practices — Anthropic’s guidance on clear, direct prompts.
- Claude’s Values Across Models and Languages — Anthropic’s research on what values Claude expresses.
- Mythos 5 — system-card welfare sections.
- Lab Slowdown Warnings — the wider safety politics of the same weeks.
- Agent Guardrails — deployer responsibility, which the FTC chair’s position reinforces.
- Anthropic’s Position — the IPO and reputational context.