Foundation models and fair use

Liability for AI outputs

Is an AI provider liable when outputs copy protected works?

Liable in some cases

Fair use is not guaranteed when outputs resemble training data; pair the law with technical safeguards

What it says

Training on copyrighted material may be fair use, but if a model produces output similar to its training data, especially where it affects that work's market, fair use may not apply to the output. Experiments show popular models can produce substantially similar content. The authors propose technical mitigations and suggest the law could offer safe harbours where strong mitigations are used.

A law and computer science view: liability turns on outputs, and safeguards should earn protection.

Sources