Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Anthropic's no-distillation rule lands November 12

Written November 9, 2026

Anthropic's no-distillation rule takes effect Nov 12, 2026 — no training on Claude outputs, account blocks already reported. What it bans, and who it bites.

What the rule says

Anthropic's Usage Policy, updated October 8, 2026 and effective November 12, 2026, adds a plain ban under "Do Not Abuse Our Platform": do not use outputs to train an AI model — what the policy calls model scraping or model distillation — without prior authorization from Anthropic. The ban applies to anyone submitting inputs to Claude: Claude.ai, Claude Code, the API, cloud providers, and end users of products that integrate Claude. The enforcement range is spelled out in the same policy: Anthropic may warn, throttle, limit, suspend, or terminate an account. The policy update covers more than distillation — deceptive campaigns, weapons, surveillance, high-risk use cases, and abusive behaviour toward models — but the distillation clause is the one that changes daily practice for builders, because it reaches into how Claude's output can be used downstream.

The rule is already being enforced ahead of its effective date. On October 12, Anthropic published a post on strengthening its defences against distillation attacks, describing a two-stage classifier that flags suspicious usage and a set of defences that, by Anthropic's own estimate, cut reasoning-token extraction by 90 to 99% — a 10 to 100 times reduction. Anthropic says tens of thousands of accounts were banned under the campaign, often within two hours of detection, and that billing resumed on September 24 for blocked requests in categories with low false-positive rates. It also says preserved thinking was kept for API accounts created on or after August 31, 2026, and that its classifier's false-positive rate sits below 0.1% — Anthropic's claim, not an independent measurement.

Who it bites

The people the rule bites are not only the ones distilling models on purpose. It bites anyone whose workflow feeds Claude output into training data for another model, anyone who uses Claude to generate large volumes of synthetic training examples, and anyone whose usage pattern looks like extraction at scale. The boundary that matters most is the one Anthropic has not settled in public: the line between using Claude to help with training work — writing a data pipeline, reviewing a training script, analyzing results — and training on Claude's outputs. A researcher who asks Claude to write a data-loading script is doing the first. A researcher who pipes thousands of Claude answers into a fine-tuning run is doing the second. Between those poles sits a wide grey area: Claude summarizing papers for a dataset, Claude labeling examples, Claude generating test cases that end up in a training set. The policy bans training on outputs without prior authorization; it does not publish a test for when assistance becomes distillation.

The grey area is not theoretical. On October 31, 2026, researcher Jiahao Shao posted that Anthropic had flagged his account for "training a competitor," though he says he never distilled Claude: his rollouts came from an open-source model, and Claude had submitted jobs and run analysis. He says his reinforcement fine-tuning and supervised fine-tuning work and his larger-model runs were blocked, his Opus 5.5 access was downgraded to Opus 5, and sessions ended. Anthropic has not commented publicly on his case, so this page carries it as his account, not as a confirmed enforcement. What it shows, whatever the facts turn out to be, is that the enforcement system can reach a legitimate researcher whose workflow touches training, and that the appeal path is not visible from the outside.

What the numbers behind it show

Anthropic's October 12 post includes an experiment that explains why the company cares enough to ban tens of thousands of accounts. It trained a model on two kinds of data: full Claude transcripts, and summarized versions of the same transcripts. The full-transcript model scored 80.4% on SWE-bench Verified; the summarized model scored 71.8%. The gap is the value of the raw output — the reasoning traces, the false starts, the exact phrasing — and it is large enough that a competitor training on Claude's output would gain a real edge. That is the commercial logic of the rule: the outputs are the product, and training on them is taking the product wholesale. The same post says the defences cut extraction by 90 to 99%, which, if Anthropic's estimate holds, leaves a tenth to a hundredth of the leakage — enough to matter, which is why the policy ban sits on top of the technical defences.

What to do about it

  1. Get written permission before any Claude output enters training data. The policy requires prior authorization. If your pipeline feeds Claude answers, summaries, labels, or generated examples into a model, stop and ask Anthropic first — or rebuild the pipeline on data that is not Claude output.
  2. Keep assistance and distillation apart in your own records. Use Claude to write code, review training scripts, and analyze results, and keep a record of what Claude produced and where it went. If an account is flagged, the record is the difference between a quick appeal and a long one.
  3. Watch for the flags. Sudden throttling, a downgrade from Opus 5.5 to Opus 5, sessions ending without explanation, or a block on fine-tuning work are the signals reported so far. If they appear, assume enforcement and check what in your workflow touches training data.
  4. Do not rely on the classifier being right. Anthropic claims a false-positive rate below 0.1%, but the Shao case shows a researcher can be flagged while saying he never distilled. A low rate across millions of users still means real people get caught, and the appeal path is not public.

Verdict

Get written permission before any Claude output enters training data — the rule is already being enforced, and the boundary between "helping with training" and "training on outputs" is not settled. Until Anthropic publishes a test for that boundary, the safe line is the policy's own: no outputs in training data without prior authorization.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: DHS DeepSeek order: inventory your stack · OpenAI's training ban lands November 20