Hacker News Reader

Comment by wodenokoto | Hacker News Reader

My personal opinion is that the source data does not exist inside the model, so the model does not in and off itself comprise of a copyright violation.

It is also not a derivative work, as it is not recognizable as any of the works it was trained on.

However, if the output it produces is close enough to existing copyrighted works, than that output cannot be used without a license.

That seems fair, and just like how we would judge a human being.

View on HN

I don't think anyone knows of the source data is encoded somehow in the model. Sure it is not explicitely, but it has been shown (at least for simpler models but also for copilot) that you can with the right prompts get them to output exact copies of results (even with possible typos in comments). That to me means the models do in fact somehow contain the source.

On the topic of derivative, backpropagation the method of training a model relies specifically on calculating the gradients or derivatives of your data.

But those gradients get averaged out, they don't stay separate for each input example.

So by averaging you remove it being derivative work?

The trouble is that these models are very specialised, and have an enormous corpus of data they have trained on. No human has read every line of open source code on GitHub, comparing what these "AI" models do to human intelligence conflates two very different arguments.

This is much clearer cut in UK law AIUI as without Fair Use we don't have the rights to train an AI. The model is fed works, you can't feed it copyright works without permission, good luck training it to make facsimile works that weren't already generic (bear in mind the Red Bus Case, however).

What I wonder here is if you train a model with no Disney works, then tell it "draw a black mouse with red shorts" and it draws a close facsimile of Mickey RTM then what? You created the work without copying and provably so; you can derive from that work without infringement ... that would seem to break copyright in rather wonderful way?!?

The Mickey issue is more about trade marks than copyright per se, and originality is not a defence to trade mark infringement.

However, there is certainly a broader spectrum of "copyright infringement" than simply cutting and pasting text verbatim.

>My personal opinion is that the source data does not exist inside the model, so the model does not in and off itself comprise of a copyright violation.

The "source data" does not exist in a ZIP file either. Neural networks are just a form of data compression like ZIP, though more opaque and lossy, but still, a highly compressed abstract version of the training corpus is encoded in the weights.

A zip file is not a good analogy to a language model as the zip can reproduce the original exactly with 100% recall, language models have very few episodes of verbatim reproduction and coverage is spotty. A zip file can only reproduce the original, but a language model can be asked to write custom, situated code. The language model still works if you block all the training data from appearing in its output.

Potential output has nothing to do with the question whether the training data is "memorized" in the weights or not. To me, the answer to the latter question is "obviously yes".

Like I said, it's a lossy form of compression.

Reply on news.ycombinator.com