A post titled "Anthropic blocks users for anything touching training larger models" reached the Hacker News front page on October 31, 2026, with about 70 points and a small comment thread. The link goes to an X post by Jiahao Shao (@jiahaoshao1), who says they never distilled Claude: rollouts came from an open-source model, and Claude only helped submit jobs and run analysis. They say they were still tagged for "training a competitor," and that anything touching RFT, SFT or training larger models now gets them blocked, downgraded from Opus 5.5 to Opus 5, or the session killed. That is one user's account. Anthropic has not confirmed it, so treat the headline claim as unverified.
This matters to anyone who builds with Claude: if a classifier flags research talk about large-model training and blocks the account, that is a practical risk for ML teams, students and tool builders. It also lands in a week of other Anthropic enforcement stories, such as the reported rule that blocks Claude Code from recreating existing video games and the Opus 5.5 reverse-engineering refusals.
TL;DR
| Question | Answer |
|---|---|
| The claim | Anthropic blocks users whose work touches training larger models |
| Source | An X post by Jiahao Shao, surfaced on Hacker News |
| Anthropic's published rule | Outputs cannot be used to train models competitive with Anthropic's |
| Confirmed block of this account? | No Anthropic statement found |
| Alternative explanation raised | Region and VPN signals tripping classifiers (commenter speculation) |
| What to do | Read the terms, avoid output-as-training-target use, keep a backup |
What does Anthropic officially say?
Anthropic's help-center article Can I use my outputs to train an AI model? says you own the Outputs you generate but cannot use them to train or develop AI models without Anthropic's written permission. It states: "Our Terms do not allow the use of Outputs to train models that are competitive with Anthropic's own."
The article lists what is allowed: training models that do not compete, such as sentiment analysis, content categorization, summarization, information extraction, semantic search and anomaly detection, as well as using Outputs to power product features, write customer-facing content, analyze data or improve internal workflows. It lists what is prohibited: training general-purpose chatbots, open-ended text generation models, using Outputs as training targets, and reverse engineering training methods. Supporting a third party's attempt at any of these also violates the terms.
Note what this policy covers. It is about using Claude's outputs as training data for competing models. It does not, in the text we read, say that asking Claude questions about how large models are trained is itself a violation. The distance between those two things is the heart of the dispute.
What do Hacker News commenters report?
The thread is small, and the comments are anecdotes, not tests. We present them as commenters' reports.
- Old news, per one commenter. nubinetwork says Anthropic has been doing this for months.
- Possible region trigger. tristanj notes the poster appears to be based in Hong Kong, where Claude is geo-blocked, and argues that doing model-training work over a VPN is a suspicious signal that could trip classifiers. This is a hypothesis, not a finding.
- Refusals for research questions. leobuskin reports the same behavior from the US and EU, with anything related to modifying existing weights or state-of-the-art training approaches, and says it is not a geolocation issue. pkoird adds a recollection that Claude was deliberately botching semi-advanced training questions.
- The opposite experience. giancarlostoro says they use Claude to train a few BERT models, and gdiamos writes that they ask Claude questions about large models like big mixture-of-experts all the time and have never had a refusal, though they avoid asking it to label training data.
The disagreement is the point: experiences range from no refusals to refusals for anything near frontier training, and none of it is controlled.
Is this new? What search turned up
One Hacker News commenter says this has been going on for months, and the help-center policy on competing models is not new. We found no first-party Anthropic statement describing how enforcement works, so we cannot say whether this is a recent change. If you need the exact current safeguards for your model, check Anthropic's own documentation, not secondary summaries.
The wider pattern is that labs increasingly police how their models are used to build rival models. We looked at a related example in our post on how Anthropic's self-assessment catches misuse, where Claude grades its own chats for misuse at a small compute cost. A classifier tuned for misuse will produce false positives, and what counts as misuse depends on policy lines that are drawn by the company.
Claimed versus verified
| Claim | Status |
|---|---|
| Anthropic prohibits using Claude outputs to train competing models | Verified in Anthropic's help center |
| Anthropic blocks users for anything touching training larger models | Unverified; a single user's report on X, no Anthropic confirmation |
| The block was triggered by training-related content | Unverified; one commenter suggests region and VPN signals instead |
| Claude refuses research questions about training | Anecdotal; commenters disagree |
| The restriction is new | Disputed; one commenter says it has gone on for months |
Why this matters for builders
If you build ML products, three practical risks follow from a policy like this, whatever the truth of this specific case.
Account risk. A classifier does not know your intent. If your day job involves model training, evaluation or data labeling, keep a second provider ready and do not run critical pipelines on a single account. This is the same argument we made in our piece on multiple AI accounts versus one big plan.
Terms risk. Using Claude outputs as training targets for a general-purpose model is explicitly barred. Distillation-style pipelines, synthetic data from Claude for a general chatbot, and labeling with Claude for such a model all sit inside the prohibited list in the help-center article. Narrow, non-competing uses such as classifiers and extraction models are explicitly allowed. If your use falls between the two, ask Anthropic for written permission rather than guessing.
Transparency risk. The most useful thing a lab can do is say plainly what triggers a block and give an appeal path. In its public letter, the company says it wants trust from users; see our reading of the Anthropic IPO letter's no-ads pledge. Whether enforcement matches that tone is something users will judge from cases like this one.
Where the line sits: output reuse versus asking questions
Three different activities get lumped together in this debate, and they carry different levels of policy risk.
1. Using Claude outputs as training data for a competing model. This is the clearest case. The help-center article names general-purpose chatbots, open-ended text generation models and "using Outputs as training targets" as barred. Distillation falls here, and so does bulk synthetic data aimed at a general model. A block for this behavior would be consistent with the published terms.
2. Using Claude to help build non-competing models. The same article allows outputs for classifiers, extraction, summarization, semantic search and anomaly detection. Commenters who say they train BERT-style models with Claude's help are describing a use the policy appears to permit.
3. Asking Claude conceptual or engineering questions about large-model training. This is the gray zone. The policy text we read does not mention it. If a classifier is built to catch category one, it can still misfire on category three, because the two share vocabulary: pretraining, data mixtures, weights, scaling. The Hacker News reports of refusals on questions about modifying weights or state-of-the-art approaches are about this third category, and they are the ones that need an Anthropic explanation.
A useful test for readers: if your prompts ask Claude to produce large volumes of text to be used as training targets, you are in category one whatever vocabulary you use. If you are asking how a technique works, you are in category three and a block would be surprising under the written policy. Neither can be resolved without Anthropic saying what its safeguards actually look at, which is why we keep the main claim in the unverified column.
How to check it yourself
- Read the help-center article linked above and Anthropic's usage policy before building a data pipeline on Claude outputs.
- If a block happens to you, note the exact message, your region, whether you used a VPN, and the last prompts, then appeal through support.
- Do not infer a cause from one account. Test with a clean account and a neutral prompt on training concepts to see whether the same refusal repeats.
- Compare with how other labs treat similar work. Our write-up of Anthropic's work on reward hacking and localized adapters shows the company is open about training research itself, which makes the question of what users may do with outputs a policy one, not a secrecy one.
What to watch next
- Whether Anthropic or the user publishes the block message and the reason given.
- Whether Anthropic clarifies if questions about training are in scope or only output reuse.
- Whether more users post reproducible cases from different regions.
- Whether the company updates the help-center article to describe enforcement.
Related reading
- Claude Code reportedly blocks recreating existing video games
- Opus 5.5 reverse-engineering refusals: claimed vs verified
- Anthropic self-assessment for misuse detection
- Dario Amodei responds to criticism
- Multiple $20 AI accounts versus one big plan
Based on the X post, the Hacker News discussion and Anthropic's help-center article as read on October 31, 2026. Nothing here is confirmed by Anthropic.