Nicholas Carlini on why privacy papers cannot be used to argue generative AI copyright issues

Machine learning researcher Nicholas Carlini writes that the starting point of his memorization research was privacy and security rather than copyright, so legal cases should not directly cite his papers as evidence. He notes that the only reliable conclusion that can be drawn from his research is that models sometimes output training data verbatim—for example, in 2020 he extracted only 600 unique training samples from the 1.5 billion parameter GPT-2, but this neither proves the regular behavior of all models nor supports the claim that "models necessarily infringe copyright."