Anthropic blocks users for anything touching training larger models

405 points · 227 comments on HN · read original →

Points and comments are a snapshot, not live.

Anthropic blocks paying users from using Claude for training or research automation.

Jiahao Shao claims Anthropic blocked him for 'training a competitor' despite never distilling Claude-rollouts came from an open-source model, Claude only helped submit jobs and run analysis. He says anything touching RFT, SFT, or training larger models triggers blocking, Opus 5.5 downgraded to Opus 5, or session killed. Shao calls it hypocritical: Anthropic trained on the world's data, then treats paying researchers doing research automation as the enemy.

In a follow-up, he argues the anti-distillation crusade is a double standard-Anthropic trained on the open web, but bans learning from its outputs at scale. Another commenter notes Anthropic trained on the open web then blocked the user who tried to train on the outputs.

What commenters are saying

Several commenters confirm similar experiences with Anthropic's safeguards: triggering on anything related to modifying weights, SOTA training approaches, or low-level hardware interactions. One notes geolocation isn't the issue-US/EU users report the same blocks. Another notes that classifiers simply degrade session quality rather than block outright, with one commenter observing it's 'a whole bunch of nothing, the guy did not get banned, he's just stuck in a bad session.'

Some argue the company's stance is hypocritical: Anthropic took the world's data without paying, then locks down its outputs. Others defend the policy as self-consistent-distillation risks losing safety controls-or suggest VPN use from Hong Kong (where Claude is geo-blocked) likely triggered the flag.