Hacker News Reader

Comment by tpmoney | Hacker News Reader

>An AI product - more specifically, the model weights - is a derivative work of the original works used for training; AI training is a process of algorithmic compression of the originals.

I'm not so sure this is as obvious a conclusion as you think. Imagine for a moment an AI OCR program. If one goes to their local library and scans all the books there to generate the models used to OCR text, does that make the OCR model and application derivative works of the books? Does copyright give Tolkien's estate the right to prevent the distribution of AI based OCR if a published copy of The Hobbit was used in creating the model? Certainly with the right inputs, the model can be used to generate a verbatim copy of the work it was trained on, but is that sufficient to say that your OCR model is just an "algorithmic compression" of these books?

View on HN

I totally agree that it's not obvious that an ML model is a derivative work. the language of the Copyright Act uses "recast, transformed, or adapted" to describe derivative works, and a pile of model weights isn't clearly that, IMO. I think it's fair to say that inferences directly replicating the creative and expressive elements (because factual information isn't copyrightable!) of a copyrighted work infringe. but I don't think it's obvious that the model itself does.

> If one goes to their local library and scans all the books there to generate the models used to OCR text, does that make the OCR model and application derivative works of the books?

there is a court case [1] addressing an even more infringing use case: scanning and OCR'ing books to produce a searchable database. that case turned on fair use, however, and not whether the database was a derivative work.

[1] https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,....

This analogy isn’t quite right, it’s more like if you trained a font generation AI using commercially licensed fonts, or trained a literature generating model on samples of copyrighted fiction.

The part that matters is that the model is being trained on the copyrighted features of the input, not the parts the copyright holder doesn’t care about.

What it's trained on shouldn't matter at all. What should matter is what it's capable of outputting, and whether that covers works in which someone holds copyright - so something regurgitating its input would be problematic, but something not capable of producing the same format of output as its input should be in the clear. Otherwise you're talking about something that shouldn't come under copyright law.

Reply on news.ycombinator.com