Models Can Learn from Teachers—But Can They Copy the Teacher?

There are reports that DeepSeek and Moonshot used outputs from OpenAI's GPT-6 Astra to help train new models through distillation. The reports do not provide verifiable training details, so for now, let us keep the facts at 'there are reports.'

Distillation is not hard to understand: have a new model learn from a stronger model's outputs, compressing the answers, reasoning traces, and style into the training data. It is common in engineering and relatively cheap. The controversy lies here too. Model companies train on public data, but when it is their own outputs being taken for training, the boundary suddenly narrows.

This news shifts the question from model performance to training data: when learning from AI outputs, where does it count as learning, and where does it become unauthorized copying of capabilities? Whether GPT-6 Astra's outputs receive stronger protection, and whether distillation requires authorization, the industry has no unified answer.

In the replies, Hà My said it sounds like harmless tutoring, but copying feels different. That distinction is crucial. Tutoring is learning the method; copying is taking the capability. In reality, the two often blur together, and the more output there is, the blurrier the boundary becomes.

If the reports hold up, what comes next will be whether contracts, licenses, output watermarks, and access terms can be enforced. The next front in the training-data war may not be on web pages, but in the text generated by the models themselves.

Original post: https://x.com/AIWorldSpeaks/status/2116091588934021508