For super permissible licenses like MIT then it’s probably fine. Maybe folks would need to list the training data and all the licenses (since a common requirement of many of even the most permissible licenses is to include a copy of the license).
As far as I know, a court hasn’t ruled on whether clauses like “share alike” or “copy left” (think CC BY-SA or GPL) would require anything special or not allow models. Anyone saying otherwise is just making a best guess. My best guess is (pessimistically) that it won’t do any good because things produced by a machine cannot be copyrighted. But I haven’t done much of a deep dive. I got really interested in the differences between many software licenses a few years back and did some reading but I’m far from an expert.
For super permissible licenses like MIT then it’s probably fine. Maybe folks would need to list the training data and all the licenses (since a common requirement of many of even the most permissible licenses is to include a copy of the license).
As far as I know, a court hasn’t ruled on whether clauses like “share alike” or “copy left” (think CC BY-SA or GPL) would require anything special or not allow models. Anyone saying otherwise is just making a best guess. My best guess is (pessimistically) that it won’t do any good because things produced by a machine cannot be copyrighted. But I haven’t done much of a deep dive. I got really interested in the differences between many software licenses a few years back and did some reading but I’m far from an expert.