Jiahao Shao says Anthropic blocked his research account on Oct. 311. He aimed the complaint straight at Dario Amodei, Anthropic's chief executive, on X. Shao is a Ph.D. student in computer science at Hong Kong University of Science and Technology. He studied automation at Zhejiang University first10.
Anthropic tagged his account for training a rival model, he wrote. He insists the actual training ran on an open-source model. Claude only submitted jobs and ran analysis, by his account1.
What Shao Says Happened
Shao's complaint named exact tools1. RFT and SFT jobs now trigger a block, he wrote. Claude Opus 5.5 drops to Opus 5 some sessions. Other sessions end outright, he said.
Shao put it bluntly on X. "You built your models on the world's data, then treat paying users doing research automation as the enemy," he wrote1. He called the policy one-sided. Anthropic, he said, already trains on everyone's writing.
Jiahao Shao@jiahaoshao1Shame on @DarioAmodei / Anthropic. I never distilled Claude—rollouts came from a strong open-source model; Claude just helped submit jobs and run analysis. Still got tagged for "training a competitor," and now anything touching RFT, SFT, or training larger models gets me blocked,…
The complaint found a wide audience fast. A Hacker News thread about the block reached 264 points within a day2.
Commenters split over who deserved blame. Some sided with Shao. Others defended the filter. A bigger threat than one session was on their minds.
One thread regular pointed to frontier labs quietly trading researchers with each other. To that reader, Anthropic's own policing looked uneven.
Anthropic's Rule Predates the Reason
Anthropic rewrote its usage policy on Oct. 84. The new version runs 4,424 words, up from 2,473, a jump of 79 percent9. Every added line bans cruelty toward Claude or fake online campaigns, a word-count study found.
Anthropic's Usage Policy Grows 1,951 Words Longer in One Rewrite
Word count of Anthropic's usage policy, the version effective Sept. 15, 2025, against the version effective Nov. 12, 2026
Each column counts the words in one version of the policy; the second column is 1,951 words longer, but the sentence banning distillation appears in both.
Source: pk-sharma.com [9]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Sept. 2025 policy | 2,473 words |
| Nov. 2026 policy | 4,424 words |
Unite.AI covered that same rewrite the same week, and named those same two subjects4. That coverage skips distillation entirely9. Anthropic barred the practice months earlier, in writing. Its February post named three Chinese labs it accused of training on Claude's answers: DeepSeek, Moonshot, and MiniMax5.
Anthropic's own words draw a line, though. "Distillation is a widely used and legitimate training method," the company wrote in February, noting that labs routinely shrink their own models the same way5. Shao says his case sits on the legitimate side of that same line1.
Why Anthropic Built the Filter
Chinese labs pushed Anthropic to act first5. Three of them ran more than 16 million exchanges with Claude through roughly 24,000 fake accounts, Anthropic said in February. The company built classifiers to catch the pattern6.
Anthropic says the defense already shows results. New safeguards cut what distillers can harvest from Claude by 90 to 99 percent, the company reported in October6. One safeguard, called preserved thinking, locks a user's past prompts, so a distiller must leave old context untouched.
Anthropic Says New Defenses Cut Distillation Harvest by Up to 99 Percent
Percent reduction in data harvested from Claude by distillers, before and after Anthropic's defenses, reported Oct. 12, 2026
The blue dot is the low end of Anthropic's own estimate, the magenta dot is the high end, and the rod spans the range the company reported.
Source: Anthropic [6]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Low estimate | High estimate |
|---|---|---|
| Data harvested by distillers | 90% | 99% |
The same safeguards carry a false positive rate under 0.1 percent, Anthropic says6. Even at that rate, the company has banned tens of thousands of accounts, often within two hours of a campaign's first move.
Tarun Chhabra leads national security policy at Anthropic. He worked on the White House National Security Council first8.
Chhabra argued the stakes reach past one company's balance sheet. Speaking at a Stanford forum in August, he said, "And without that distillation, we could probably have a lead closer to around 18 months"8. He meant the gap between American and Chinese frontier models, a gap that narrows, he said, every time a rival lab copies a shortcut.
The Case Against the Net
Some researchers reject the national security framing entirely. Nathan Lambert, post-training lead at the Allen Institute for AI, called the broader crackdown a mistake back in May7. "Associating all of distillation with these attacks, which is to date an industry standard for post-training, from open and closed models alike will be a massive own goal," he wrote.
Lambert drew a sharper line in the same post. "What these few labs are doing should be referred to as jailbreaking or abuse, rather than distillation," he wrote7. He warned that lumping the two together risks the open model ecosystem that researchers like Shao depend on.
Shao's account fits the distinction Lambert draws. He insists an open-source model did the learning, while Claude only helped manage the job1. Anthropic's own numbers admit a filter this size still misses a rare case now and then6. Shao says his case fell inside that rare margin.
What to Do
Test any research pipeline that touches reinforcement or supervised fine-tuning against Claude's current rules before a deadline forces the question. Save every prompt and response tied to a blocked session. A saved record speeds up an appeal to Anthropic's support team.
Separate your training data sources in writing before a project starts. Note which model generated which output, and keep that log current. The habit costs little next to a frozen account mid-deadline. A paying lab customer with a clean log reaches a human reviewer faster than one arguing from memory.
By the Numbers
- 1,951 words: how much longer Anthropic's rewritten usage policy runs than the one it replaced9.
- 16 million: exchanges Anthropic says three Chinese labs ran with Claude before its February crackdown5.
- 24,000: the fake accounts Anthropic says those same three labs used5.
- 90 to 99 percent: Anthropic's claimed cut in distillation harvest from its newest defenses6.
- Under 0.1 percent: that same filter's reported false positive rate, by Anthropic's own count6.
- 18 months: the lead over Chinese labs that Chhabra says distillation controls would buy8.
- Nov. 12: when Anthropic's rewritten policy takes effect3.
What to Watch
Watch whether Anthropic answers Shao's complaint in public. A reply would show how the company tells a legitimate researcher from a copier.
The next open-weight model that claims a jump past its training budget deserves the same scrutiny. DeepSeek faced that scrutiny in February, and the next lab to claim a leap should face it too5. A verification path for researchers would settle more of these cases than another ban.
Sources
- Jiahao Shao (@jiahaoshao1), post on X, Oct. 31, 2026, https://x.com/jiahaoshao1/status/2116475922190922088
- "Anthropic blocks users for anything touching training larger models," Hacker News, Oct. 31, 2026, https://news.ycombinator.com/item?id=50319358
- Anthropic, "Usage Policy," effective Nov. 12, 2026, https://www.anthropic.com/legal/aup
- Mira Kellan, "Anthropic Updates Usage Policy, Adding Model Abuse and Deception Rules," Unite.AI, Oct. 8, 2026, https://www.unite.ai/anthropic-updates-usage-policy-adding-model-abuse-and-deception-rules/
- Anthropic, "Detecting and preventing distillation attacks," Feb. 23, 2026, https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks
- Anthropic, "Strengthening our defenses against distillation attacks," Oct. 12, 2026, https://www.anthropic.com/news/distillation-defenses
- Nathan Lambert, "The distillation panic," Interconnects AI, May 4, 2026, https://www.interconnects.ai/p/the-distillation-panic
- "Anthropic, OpenAI, and the New Frontier of Artificial Intelligence in National Security," Stanford Freeman Spogli Institute, Aug. 5, 2026, https://fsi.stanford.edu/news/anthropic-openai-and-new-frontier-artificial-intelligence-national-security
- Parminder Kumar Sharma, "Anthropic's new Usage Policy is 79% longer, effective 12 November," pk-sharma.com, Oct. 9, 2026, https://www.pk-sharma.com/briefing/anthropic-usage-policy-79-percent-longer-effective-12-november-35-days-after-posting
- Jiahao Shao, personal website, accessed Nov. 1, 2026, https://jiahao-shao1.github.io/
