So. Yes, technically possible. But impossible by accident. Furthermore when you make this argument you reveal that you don't understand how these models work. They do not simply compress all the data they were trained on into a tiny storable version. They are effectively multiplication matrices that allow math to be done to predict the most likely next token (read: 2-3 Unicode characters) given some input.
So the model does not "contain" code. It "contains" a way of doing calculations for predicting what text comes next.
Finally, let's say that it is possible that the model does spit out not entire works, but a handful of lines of code that appear in some codebase.
This does not constitute copyright infringement, as the lines in question a) represent a tiny portion of the whole work (and copyright only protecst against the reduplication of whole works or siginficant portions of the work), and B) there are a limited number of ways to accomplish a certain function and it is not only possible but inevitable that two devs working independently could arrive at the same implementation. Therefore using an identical implementation (which is what this case would be) of a part of a work is no more illegal than the use of a certain chord progression or melodic phrasing or drum rhythm. Courts have ruled about this thoroughly.
Like this?
Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book - https://news.ycombinator.com/context?id=44972296 - 67 days ago (313 comments)
I comprehend it just fine, I was adding context for those who may not comprehend.