Hacker News Reader

Comment by DowsingSpoon | Hacker News Reader

>you can train an AI on copyrighted material just like people can learn from copyrighted material.

No amount of whining and hand wringing from engineers will ever make this true. This is for the courts to decide.

A reasonable interpretation, in my eyes, is that the training process is a black box which takes in copyrighted works and produces a training model. The training model is a derivative work of the inputs. It therefore violates the copyrights of a large number of rights holders. The outputs of the model are derivative works which also violate copyright.

And anyone using or training a model trained on works for which they do not have the rights? Completely fucked. Or at least, they must accept this as a real risk.

View on HN

> No amount of whining and hand wringing from engineers will ever make this true

Please omit flamebait and swipes, as the site guidelines ask: https://news.ycombinator.com/newsguidelines.html. Your comment would have been fine without that bit.

Derivative works are typically things like language translations or film adaptions of a novel. A large language model is something like a probability breakdown of the order of word fragments in a body of text. It's a collection of statistics and math. It's different.

Now, can you get it to output a derivative work? Maybe. Is every output a derivative work? Maybe not.

Things like sports statistics and directories of phone numbers are not copyrightable. Maybe models fall into a similar category. Could be. It’ll be interesting to see how this shakes out over the next few years.

I think a reasonable interpretation is also that what you are saying is correct, that doing all that does indeed infringe others' copyright, but that a fair use defense is valid.

I won't be particularly thrilled if that turns out to be the case, but I wouldn't be surprised if it does.

But as you say, we won't know until it's tested in court. And even then, often court cases around a complex topic like this will end up with a ruling that only clarifies a narrow aspect of it. So it might take many related court cases before we have a pretty good understanding of where the law stands. And then, of course, the law could change.

Its unmistakably not a derivative work of the inputs individually or collectively, since a derivative work must be itself an distinct work of authorship (the same as the work of authorship requirement for copyright), and the output of a purely mechanical process is not.

The collection of inputs itself might be a derivative work of the individual inputs, before considering Fair Use.

> A reasonable interpretation, in my eyes, is that the training process is a black box which takes in copyrighted works and produces a training model. The training model is a derivative work of the inputs. It therefore violates the copyrights of a large number of rights holders. The outputs of the model are derivative works which also violate copyright.

What is the blackbox “limit” here? Is the mean value of all images in imagenet (which contains many copyrighted images) violating copyright? Is the character count of sarah silverman’s books? What about a prime number representing them - https://en.wikipedia.org/wiki/Illegal_number?wprov=sfti1

Training is much more similar to a character count than an illegal prime in my view, and thus, is almost certainly going to be okay/found to be okay. If not, something like, 90% of all models used today had some component trained on copyrighted data of some form.

Reply on news.ycombinator.com