> AI being trained on copyrighted data is considered fair use because it transforms the underlying data rather than distribute it as is.

It's not a binary. Sometimes it fully reproduces works in violation of copyright and other times it modifies it just enough to avoid claims against it's output. Using AI and just _assuming_ it would never lead you to a copyright violation is foolish.

> uses same energy as more than 100 GPT questions.

Are you including training costs or just query costs?

> But the reasoning is motivated from a wrong place.

That does not matter. What matters is if the outcome is improved in the way they predict. This is actually measurable.

>That does not matter. What matters is if the outcome is improved in the way they predict. This is actually measurable.

Ok lets discuss facts.

>It's not a binary. Sometimes it fully reproduces works in violation of copyright and other times it modifies it just enough to avoid claims against it's output. Using AI and just _assuming_ it would never lead you to a copyright violation is foolish.

In the Anthropic case the Judge ruled that AI training is transformative. It is not binary as you said but I'm criticising what appears as binary from the original policy. When the court ruling itself has shown that it is not violation of copyright, it is reasonable to criticise it now although I acknowledge the post was written before the ruling.

>Are you including training costs or just query costs?

The training costs are very very small because they are amortised over all the queries. I think training accounts around .001% to .1% of each query depending on how many training runs are done over a year.