I think until now this has been a complaint about generative AI applications that has been handwaved away ("what if co-pilot generates GPL code?" -> "Well it wi

I think until now this has been a complaint about generative AI applications that has been handwaved away ("what if co-pilot generates GPL code?" -> "Well it will generate a blend of many pieces of code") If there is now evidence that it does memorize and emit training data, that argument had new life. That said, I imagine it's easy enough to add a post-process step to ensure your result isn't a member of the training set.