The Importance of Copyright in AI Systems

The other place aside from training is outputs of models those can definitely violate copyright right so I think so sometimes we do reproduce works that are more or less verbatim copies of what we trained on. Subsequently like OpenAI is mitigated against that so you can't do it but then Daphne's had a bunch of follow-on work showing that because of the number of parameters in these large models including like the text-to-image models you can memorize items in your training set. There's something called the substantive similarity test that might it might run afoul of so Outputs can be indefinitely generated violate copyright.