Claude switched models during AI training work? A PhD student accuses Anthropic

Jiahao Shao, a computer science PhD student in Hong Kong, wrote on X on October 31 that Anthropic flagged him for “training a competitor” and that since then Claude has switched models or blocked everything that touches RFT, SFT, or training larger models. His post reached the front page of Hacker News, where it collected over 250 points and more than 130 comments.

What actually happened?

Claude switching models or blocking a task during AI work is a documented safeguard. Whether Shao was caught by it fairly, we only know from his own account so far.

Shao writes that he never distilled Claude. According to him, the training “rollouts” came from a strong open-source model, and Claude only helped submit jobs and analyze results. Even so, he got the “training a competitor” label, and his post opens with “Shame on @DarioAmodei / Anthropic”.

At the time of writing, Anthropic has not commented on the case publicly. Nobody from the company has spoken up in the Hacker News thread either.

Why does Claude switch models?

Here we have solid sources. The Claude Opus 5.5 announcement says it is the first Opus model with safeguards similar to those in Fable 5.1, and that blocked biology and so-called “frontier LLM development” tasks were handled by Claude Opus 5 in testing. The details are in a Claude help center article. There are four categories:

1. Cybersecurity: offensive tasks (e.g. writing exploits, penetration testing) move from Opus 5 or 5.5 to Opus 4.8. Secure coding and scanning your own code still work.
2. Biology: for “dual-use” tasks (virology, toxicology, molecular design), Opus 5.5 switches to Opus 5.
3. Frontier LLM development (Opus 5.5 only): a narrow set of skills, such as writing kernels for certain ML accelerators, also ends with a switch to Opus 5. According to the article, this shouldn’t affect most AI and ML work.
4. Distillation: attempts to extract the model’s internal reasoning are blocked outright, with no switch to another model.

On top of that, Anthropic’s usage policy prohibits using Claude’s outputs to train AI models (it mentions “model scraping” and “model distillation”) without the company’s permission. The current version takes effect on November 12, 2026.

Why is this getting so much attention?

Because Shao is a researcher, not an anonymous account farm. As one commenter puts it, researching how to build better AI is still research, not commercial activity.

The second reason: other users describe similar situations. In the comments, one person writes that Fable refused to look through a backup folder and wanted to switch to an older Opus, another that Claude Code wouldn’t save a task handoff file because it saw “extraction” in it. Several people also complain about overzealous classifiers when working on hypervisors and drivers.

The third reason: this is another case within a few days where Claude’s rules surprise users, after the block on game clones in Claude Code.

Is something wrong here?

First, we only know one side. We don’t know what queries Shao actually sent or what the “label” he writes about looks like. One commenter suspects that using a VPN in Hong Kong played a role, another replies that the safeguards don’t check IP addresses. Let’s treat both theories with caution.

Second, the official description of the safeguards is narrower than what Shao describes. According to the help center, the “frontier LLM development” safeguard covers “a small set of capabilities”, and distillation means extracting the model’s reasoning. Blocking everything that touches SFT or RFT would be much broader than what the company states, and that probably calls for an explanation from Anthropic.

Third, some of these refusals are billed. According to the help center, refusals before a response is generated for biology, distillation, and “frontier LLM development” are charged, and the fallback model’s response is billed at its own rates.

What should you watch out for?

If you work on training models with Claude’s help, it’s worth knowing a few options from the help center:

1. After a model switch, edit your previous message before going back to Opus 5.5, because the same safeguards may trigger again.
2. In Settings > Capabilities you can turn off “Switch models when a message is flagged”. Then, instead of a switch, the conversation pauses and you decide what to do next.
3. Report mistaken blocks via “Send feedback”.
4. In the API, automatic switching is not on by default; you configure fallbacks yourself.

(There is a verification program for cybersecurity and life sciences; the help center doesn’t mention one for ML researchers.)

FAQ

Why did Claude switch models in my conversation?

Because the Opus 5 or 5.5 classifiers flagged the request as risky (cybersecurity, biology, or “frontier LLM development”), and Claude ran it again on a less capable model. The conversation then shows a notice about the change.

Can Claude help with training AI models?

According to the help center, the safeguard shouldn’t affect most AI and ML work. What’s prohibited is training models on Claude’s outputs without Anthropic’s permission, and narrow categories, such as kernels for ML accelerators, switch to Opus 5.

Can you turn off model switching in Claude?

Yes, in Settings > Capabilities with the “Switch models when a message is flagged” option. A blocked request then won’t move to another model; the conversation simply pauses.