When does generative AI qualify for fair use?

Summary

When does generative AI qualify for fair use? is an essay^[1].

Key Facts

• When does generative AI qualify for fair use? authored Suchir Balaji^[2].
• When does generative AI qualify for fair use?'s instance of is recorded as essay^[3].
• When does generative AI qualify for fair use?'s language of work or name is recorded as English^[4].
• When does generative AI qualify for fair use?'s publication date is recorded as +2024-10-23T00:00:00Z^[5].
• When does generative AI qualify for fair use?'s main subject is recorded as generative artificial intelligence^[6].
• When does generative AI qualify for fair use?'s main subject is recorded as fair use^[7].
• When does generative AI qualify for fair use?'s work available at URL is recorded as https://suchir.net/fair_use.html^[8].
• When does generative AI qualify for fair use?'s title is recorded as When does generative AI qualify for fair use?^[9].
• When does generative AI qualify for fair use?'s first line is recorded as While generative models rarely produce outputs that are substantially similar to any of their training inputs, the process of training a generative model involves making copies of copyrighted data.^[10].
• When does generative AI qualify for fair use?'s last line is recorded as That being said, none of the arguments here are fundamentally specific to ChatGPT either, and similar arguments could be made for many generative AI products in a wide variety of domains.^[11].
• When does generative AI qualify for fair use?'s copyright status is recorded as copyrighted^[12].
• When does generative AI qualify for fair use?'s quotation or excerpt is recorded as The basic intuition behind this bound is that low-entropy model outputs are more likely to be including information from the model’s training data. In the extreme case, this is the problem of regurgitation, where a model deterministically outputs parts of its training data. But even nondeterministic samples can still use information from the training data to some degree -- the information may just be mixed in throughout the sample instead of directly copied.Note that there’s no fundamental reason why the entropy of a model’s outputs needs to be lower than the true entropy, but in practice model developers tend to choose training and deployment procedures that favor low-entropy outputs. The basic reason for this is that high-entropy outputs involve more randomness in their sampling process, which can cause them to be incoherent or contain hallucinations.^[13].

Body

Works and Contributions

When does generative AI qualify for fair use? authored Suchir Balaji^[2].