Copyright safety for generative AI

AI training and copyright

Can AI be trained on protected works?

Closer to freely allowed

Training is generally a lawful non-expressive use; the risk lies in memorised outputs

What it says

Sag places generative AI in the line of fair use cases on copy-reliant technology such as Google Books and HathiTrust. Models generally learn abstract features rather than copying expression, so training is usually non-infringing, but memorisation is a real risk in some settings, and he proposes copyright safety guidelines to reduce it.

The standard US account of why training can be fair use, and where it stops.

Sources