Hacker News Reader

Hacker News

Comment by kmoser | Hacker News Reader

Full thread

kmoser

Lots to comment on but this stood out:

> “A lawsuit like this heightens the demand for Generative AI replacements.”

Most generative AI corpora were arguably trained on copyrighted material, making the output potentially infringing.

View on HN

> Most generative AI corpora were arguably trained on copyrighted material, making the output potentially infringing.

Training is not neccesarily sufficient for it to be a derrivative work, just like if you learned to draw based on famous drawings doesn't mean every single drawing you ever made is infringing.

Obviously there are cases where it could be infringing, its going to depend how close the output is to the original.

I guess it depends on how you read the post, is it saying use gen-AI to intentionally recreate the photo, something that sounds danger-zone, or are they saying use gen-ai to make some other photo suitable for purpose?

I'm largely out of this space now but my understanding is that some copyright cases around model training are winding through courts but I haven't seen anything definitive come out. The IP lawyers I know are skeptical but we'll see.

EU AI Act is moving towards genAI output being non-copyrightable and that you'd need to actually prove derivative character from a specific copyrighted work(s) to claim infringement.

AFAIK american law is going towards similar setup.

IANAL but, yes, with US/UK (i.e. common law regimes) that's something along my understanding as well. Which I generally agree with even if some/many readers here probably do not. Of course, output being copyrightable and copyright infringement on the inputs are two different things.

An important point in copyright infringement is that it generally applies on distribution to other parties.

So the process of acquiring inputs may or may not be an infringement, but with at least proposed EU rules it does not matter to created model itself.

The exception being that output it produces is judged similar to infringement as human output without any "transformative work" credit to model - so similar to how a human could learn a book or painting to memory and close enough reproduction from memory would be infringement, but not generally using the ideas taken from them

Actually, copyright generally is infringed when a copy is made; hence the name.

That's why, say, 17 USC 106 lists reproduction as the first exclusive right of a copyright holder. And why Berne Article 9 [^1] is about restricting right of reproduction to the author.

Damages are often, in different jurisdictions, related to actual harm. So, distribution is the focus of lawsuits because actual harm in the making of a copy is usually negligible. Few people are suing to stop copying, they're suing to be recompensed for the [potential] commercial benefit derived from the copying.

In as far as you need to make a copy to use it to process and adjust the weights of an ML model, then yes this activity is an infringement to the right to control reproduction.

One of the measures for transformative use is whether the production of the copy commercially harms the original creator/author. I can't see how you can argue that ML models don't do that. Besides which we don't have an equivalent precedent to 'transformative use' in UK so where our courts can go with all this is not clear.

https://www.wipo.int/wipolex/en/text/283698

There is however the aspect of how much a work is derivative of another work, and that's where the ultimate mixing aspect of AI models shines - as it was documented that removing or adding a specific work in some cases impacted around ~1bit - or less than the random number generator used in some of the processes.

This is also why the output of AI model can infringe - because while having copyrighted work in training set does not result in model being considered derivative, it can produce a work that is obviously a derivative - and as such can be claimed to be a derivative that harms original creator.

Just the fact that the model in general might impact some very vague notions of commercial income is not enough, it must do so through a derivative, and thus infringing, copy

Bartz v Anthropic is some good authority on fair use (https://storage.courtlistener.com/recap/gov.uscourts.cand.43...). Still, it can't be said to be definitive because the plaintiff's arguments on market harm (with respect to fair use, not piracy) were limited and there were, as far as I can remember, no compelling examples provided of model output reproducing large swathes of training text.

Sometimes human writers sit down to write and accidentally end up verbatim reproducing an NYT paywalled article, too, and no one bats an eye, but AI does it and allll of a sudden we’re in court? Poppycock!

> Training is not neccesarily sufficient for it to be a derrivative work, just like if you learned to draw based on famous drawings doesn't mean every single drawing you ever made is infringing.

We don't know that model training is the same thing as inspiration. Training is a mathematical process with theoretically deterministic outputs. It's converging the weights towards being able to exactly reproduce the training data, rather than parts of the training data subjectively influencing a creative output. We will just have to see how this plays in court.

There is plenty of precedent being written here. It does not seem to be the case at all for the average use of this technology.

https://www.congress.gov/crs-product/LSB10922

Ye olde double edged sword

On one hand aggressively punitive copyright claims stifle creativity and innovation in transformative art. On the other hand, generative AI reopens that transformative creativity.

Reminder that the original copyright term was 14 years, with a single 14-year extension.

If this were still the norm, it would feel crazy that blockbuster movie studios are still recycling comic book characters from the 1950s.

Agreed.

If copyright was shorter blockbuster movies could still recycle comic book characters from the 50s. It just means that any filmmaker could do it rather than just who can afford to pay Marvel/DC

Even if the specific image being infringed were not in the corpus, it's possible that a court would return a judgment of copyright infringement.

Consider the case where someone deliberately prompts the AI to build a facsimile image and the AI does a creditable job after some tweaking.

same goes for anything you output :)

Except everyone who has tried to argue that in court has lost.

Has this been argued? I'd love to read some actual court decisions.

Here are some cases (mined from Wikipedia sources):

Tremblay v. OpenAI, Inc., No. 3:23-cv-03223 (N.D. Cal.) (https://dockets.justia.com/docket/california/candce/3:2023cv...)

Andersen v. Stability AI Ltd., No. 3:23-cv-00201 (N.D. Cal.) (https://law.justia.com/cases/federal/district-courts/califor...)

Authors Guild v. OpenAI, Inc., No. 1:23-cv-08292 (S.D.N.Y.) (https://law.justia.com/cases/federal/district-courts/new-yor...)

Getty Images (US), Inc. v. Stability AI, Inc., No. 1:23-cv-00135 (D. Del.) (https://dockets.justia.com/docket/delaware/dedce/1:2023cv001...)

The New York Times Co. v. Microsoft Corp., No. 1:23-cv-11195 (S.D.N.Y.) (https://law.justia.com/cases/federal/district-courts/new-yor...)

Richard Kadrey et al. v. Meta (https://www.documentcloud.org/documents/25984135-richard-kad...)

Bartz v. Anthropic (https://www.documentcloud.org/documents/25982181-authors-v-a...)

---