🔥 Trending on HN

What Does Anthropic, the Company Behind Claude, Restrict in AI Research?

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Anthropic(an-THROP-ik)

The company that makes Claude.

RFT(R-F-T)

A method for improving an AI with scored answers.

SFT(S-F-T)

A method for improving an AI with example answers.

What happened

On October 31, 2026, X user Jiahao Shao said Anthropic, the company behind Claude, flagged him for “training a competitor.” He said he never distilled Claude. He said a strong open-source model produced the rollouts. He said Claude only submitted jobs and ran analysis. He also said work touching RFT, SFT, or larger-model training was blocked. He described Opus 5.5 being downgraded to Opus 5, or sessions being killed. These are claims in his public post. They are not an independent finding or a public case confirmation from Anthropic.

What Anthropic’s public rule says

Anthropic’s Help Center says users own Outputs made from their Inputs. It also says customers may not use Anthropic’s services to train or develop AI models without written permission. Its rule is especially clear about models that compete with Anthropic’s own models. It lists general-purpose chatbots, open-ended text generation, using Outputs as training targets, and reverse engineering training methods as prohibited examples. It says customers may train non-competing tools, including classifiers, summarizers, information-extraction systems, semantic search, and anomaly detection. Anthropic gives two reasons. Safety controls may not survive when Outputs train a new model. Anthropic also cannot see how that model will be deployed. The official explanation

The key boundary

The public complaint and the public rule describe different things. Shao’s account says Claude helped run a research workflow. It says another open-source model supplied the training material. Anthropic’s rule focuses on using its service or Outputs to build competing models. The public record does not show how Anthropic interpreted the account, or what signal triggered enforcement. That gap matters. A researcher may view job submission and analysis as ordinary automation. A provider may view the same pipeline as support for model training. That is an inference from the two accounts, not a confirmed description of this case.

What is confirmed

We can separate three facts. A public post made the blocking claim. Anthropic has a published restriction on training competing models. The supplied cluster snapshot lists 470 points and 260 comments on Hacker News. The Hacker News thread shows community attention. Points and comments do not prove the original claim.

What is still unknown

We do not have a public, case-specific explanation from Anthropic in the pages checked. We do not know whether Claude Outputs entered training targets. We do not know what exact action triggered the restriction. We also do not know whether the reported block, downgrade, and killed sessions happened together, separately, or only as possible outcomes. The post does not establish how often similar cases occur.

What to watch next

The next useful evidence would be an Anthropic explanation, an appeal path, or more documented cases. The most important clarification would distinguish using Claude’s Outputs as training data from using Claude to submit jobs or analyze results. Until then, the careful headline is not “Anthropic blocks everyone who studies large AI.” It is “one user says Anthropic restricted work touching model training, while Anthropic’s public rules limit use for competing-model development.”

💬 Claude refusals around large-model training

The thread mixes user reports that Claude blocks or interrupts work near large-model training with counterexamples involving small research and code assistance. Location, task size, and temporary classifier behavior are all debated, and the comments do not establish Anthropic’s exact policy.

  • Several users report refusals around training large models, changing existing model weights, state-of-the-art methods, or generating and classifying training data. These are user reports; the thread does not prove a universal block on all such work.
  • One commenter suggests that a Hong Kong geo-block combined with VPN use may have triggered detection. Users in the US and Europe report similar reactions, so location alone is not established as the cause; both sides are anecdotal claims.
  • A user self-report says a model using a novel linear-attention mechanism trained for at least a week without refusals. Another user says Claude only wrote code and supervised training, rather than generating training data directly, and describes the work as small-scale research. This suggests task and scale may matter, but it does not establish a rule.
  • Another user self-report says training small BERT models did not trigger safeguards because the models were small. No size threshold is provided.
  • A separate user self-report says reinforcement-learning work for a game triggered safeguards several times months earlier. Other commenters describe the incident as a safeguard interruption or bad session rather than a ban, and say a new session or rephrasing can help. None of these claims is independently verified.
  • Commenters speculate about classifiers, trust scores, and policy tiers, but the thread provides no evidence identifying the actual mechanism.
  • The ethics debate goes both ways: critics allege that Anthropic benefits from copyrighted data while restricting others’ distillation or training, while defenders say intellectual property should be protected and another party’s alleged violation does not justify a new one. The related legal and factual claims are not verified here.
  • Some users recommend alternatives, open weights, and portable workflows to reduce vendor lock-in. Others report that the strongest hosted models still make a meaningful difference on complex work; that is subjective user experience, not an independent benchmark.

initial digest at 260 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Claude, the AI helper, and Anthropic’s training rules

📰 Full story: What Does Anthropic, the Company Behind Claude, Restrict in AI Research?

One user says his research work was restricted. Anthropic’s public rule explains one reason this boundary exists.

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
RFT(R-F-T)

A way to train an AI by scoring its answers.

SFT(S-F-T)

A way to train an AI with good examples.

Outputs(OUT-puts)

The answers produced by Claude.

💡 The gist

  • X user Jiahao Shao says Anthropic stopped some AI research work.
  • Anthropic restricts using Claude Outputs to train competing AI models.
  • Hacker News showed attention, not proof, with 470 points and 260 comments.

What did the user say?

Jiahao Shao posted that Anthropic, the company behind Claude, tagged him for “training a competitor.” He says he did not use Claude’s answers to train another AI. He says a strong open-source model made the training results. Claude only helped submit jobs and study results, he says. He also says work involving RFT, SFT, or larger models was blocked. RFT and SFT are two ways to train an AI. He reports a downgrade from Opus 5.5 to Opus 5. He also reports ended sessions. Read the original post. These details come from his account. They are not confirmed by Anthropic.

What does Anthropic’s rule say?

Anthropic’s official Help Center says users own their Claude Outputs, or answers. It also says customers need written permission to use its services for AI training or development. It does not allow Outputs to train models that compete with Anthropic. It names general chatbots and open-ended writing models as prohibited examples. It also bans using Outputs as training targets. The company says safety protections may disappear in a new model. It also says Anthropic cannot watch that model after deployment. It allows some non-competing tools. Examples include classifiers, summarizers, search tools, and anomaly detectors. Read the official explanation.

Why is this hard to judge?

“Claude helped with research” can mean several things. It might mean sending jobs or comparing results. It might also be part of a larger training pipeline. The public material does not show what Anthropic detected here. It does not show whether Claude’s answers became training data. So the complaint and the rule should not be treated as identical.

Why mention Hacker News?

The story reached Hacker News. The supplied snapshot lists 470 points and 260 comments. Those numbers show attention from that community. They do not show that the complaint is true. See the thread.

What should we watch?

Look for Anthropic’s explanation of this account. Look for a clear appeal process. Look for other documented cases. Researchers may need to separate Claude’s answers from their training data. They may also need backup tools for important work. The careful summary is simple. One user reported a restriction. Anthropic has a rule against training competing models. The link between those facts remains unproven.

💬 Is Claude blocking large-AI work?

Users say Claude sometimes stops requests about building large AI models, but they disagree about why it happens and how broad the rule is.

  • Users report refusals around training large models, changing model internals, or making training data. This does not prove that every user is universally blocked.
  • A Hong Kong-and-VPN explanation is disputed by users in the US and Europe who report similar behavior.
  • Self-reported examples say small BERT training and code-only help worked, while some reinforcement-learning work was flagged. The thread suggests that size and the kind of help requested may matter, but gives no clear rule.
  • Some users say the event was a temporary safety interruption, not an account ban, and that a new session or rewording helped. This is unverified.
  • People disagree about fairness: some call Anthropic’s data practices hypocritical, while others say intellectual property should be protected. Many advise keeping other models or open-weight options because relying on one provider is risky, although some users say the best hosted models still help more with complex work.

initial digest at 260 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Did Claude, an AI helper, stop helping with a big project?

📰 Full story: What Does Anthropic, the Company Behind Claude, Restrict in AI Research?

One person says Claude stopped helping. The company has not publicly confirmed his story.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Anthropic(an-THROP-ik)

The company that makes Claude.

Claude(KLAWD)

A computer helper that answers questions.

Hacker News(HACK-er news)

A website where people discuss technology.

What happened

Jiahao Shao wrote about Anthropic, the company that makes Claude. Claude is a computer helper that answers questions. Shao said Claude stopped helping with his big AI work. He said another AI made the training answers. He said Claude only sent jobs and checked results. He said some sessions stopped or changed models. This is Shao’s story, not a company confirmation. Read the post.

What Anthropic says

Anthropic has a rule about Claude’s answers. Those answers cannot teach a competing AI without permission. Anthropic says safety protections might disappear. It also cannot watch the new AI everywhere. Read the rule.

Why people noticed

The story appeared on Hacker News, a tech news site. The supplied snapshot lists 470 points and 260 comments. Those numbers show that many people noticed it. They do not prove the story is true. See the thread.

What we still need

We need Anthropic’s explanation. We need details about the blocked work. We need to know whether Claude’s answers trained another AI. Until then, say one person reported a block.

💬 Will Claude help make a big AI?

Some people say Claude stops when asked to help make a big AI. But nobody in the thread proves the exact rule.

  • Some users say big-AI training was stopped. Other users say small AI projects or code help worked.
  • Some people blame Hong Kong or a VPN. Other users say the same thing happened in different countries.
  • Some users think it was only a temporary safety check, not a ban. A new chat sometimes helped, but this was not independently checked.
  • People disagree about fairness and protecting creative work. They also say it is safer to have more than one AI, even though some users feel the strongest AI is still better for very hard tasks.

initial digest at 260 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.

Sources