If the model is learning the nature of code and synthesizing new code then it may not violate some licenses. When people do this its usually fine. However, if

If the model is learning the nature of code and synthesizing new code then it may not violate some licenses. When people do this its usually fine. However, if the model is simply compressing code it sees into the model weights, just memorizing snippets, then outputting those snippets, that’s much more likely to violate licenses. Like when people copy+paste from some licensed code without attribution (even if realistically nothing is enforced most of the time). The truth is likely somewhere in the middle. Lets say 20% of code is stored directly in the weights but the rest is synthesized. That’s a problem for the whole product. We already know models do a little bit of both depending on the data coverage. Common structures like if/then, loops, etc. are probably “understood” because the model saw them in lots of contexts. However, specific functions, especially those that are seen only a few times in much the same contexts are more likely to be copied. There’s a spectrum here from shortcut learning to understanding. OSS doesn’t really have the access or resources to evaluate this. Github isn’t really incentivized to share any analysis they’ve done here. What’s interesting to me is that their solution to this problem is to put the issue on their users/customers. By default, crawl everything public, ignorant of licenses, and if the customer has license concerns its on them to disable public code matching.