Is training AI models on a rival model's outputs fair competition or copying?
The Information, cited by Reuters, reported that DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra to train their latest models after Anthropic said new defenses cut distillation of Claude by about 90 to 99 percent. The companies had not commented at the time of the reports.
Side A
Distillation is copying
This camp argues that training on a rival model's outputs amounts to copying. Posters described the reported use of GPT-6 Astra outputs as copying homework at scale and said it undercuts claims that the models were built from scratch. One said cheap open models are still useful but that labs relying on distillation should not be called independent frontier labs.
Who took this side
No public figures have taken this side on X yet
Evidence posts
Overlap
Where they overlap
No clear common ground was found in the posts reviewed.
Side B
Distillation is fair game
This camp argues that distillation is legitimate competition or that objections to it are inconsistent. Posters said frontier labs themselves trained on text nobody licensed, so outrage over others training on their outputs is hard to justify. One said accessible data is fair game because that is how large language models were built, and another said the only practical defense left for OpenAI is staying ahead.
Who took this side
No public figures have taken this side on X yet
Evidence posts
Reader vote
Be the first to weigh inWhere do you land? Sign in to vote — one account, one vote.
Free newsletter
Both Sides Daily
One email a day with the day's debates: what each side says, where they meet, and how readers voted. No spam, unsubscribe with one click.
This page was compiled by AI and published after automated checks. Something wrong?Report a correction