Hacker News Reader

Comment by phoehne | Hacker News Reader

I think this will be a bigger issue than some people think. Maybe there's a market for 'clean' training data that doesn't include potential copyright claims. Just public domain works. We'll know it's an AI because it talks like a late 18th century/early 19th century writer?

View on HN

This isn't completely new, similar issues came up with search engines and this may be seen as 'transformative'. But there may be issues with models that happily reproduce copyrighted texts in their entirety along with other novel issues like models that hallucinate defamatory things or other such problems.

Still, I doubt this particular genie can be stuffed back into the bottle, so we'll probably see a lot of litigation and work on alignment, etc. along with new types of abuse.

I agree it's not an entirely new issue. But it's a little different from search results. Say I use the generative paint brush in photoshop. It reproduces a portion of the copyrighted work. I then use the image on an advertising campaign, other merchandise, or post the final product as my own work. Would I be responsible? Would Adobe? Given that retraining these models is not simple, or cheap, would this be just 'cost of doing business?' Would I be able to buy insurance for this?

Enquiring minds want to know.

How the copyrighted work is reproduced is irrelevant wrt whether copyright is violated.

I suspect you would be held liable, though you would probably have a claim of your own to make against Adobe depending on the nature of the work in question.

"Errors and Ommisions" is a fairly standard name for the type of insurance you are thinking of. Typically, you would get it to cover you / your small business in the event that you, by mistake or minor negligence, caused harm to a client (i.e. a bug that cost them some sales).

I don't know the ins and outs of the insurance too well, but as long as it didn't create something super famous like the Nike swoosh, a genuine mistake through the use of an industry standard tool like Adobe might be covered.

This already sort of happened in 2008 when Chuck Close forced someone to stop distributing a Photoshop plug-in that imitated his style.

https://hyperallergic.com/54104/my-chuck-close-problem/

Hmm, from the link you provided, it was a website named freechuckcloseart.com that would autogenerate artwork in the style of chuck close.

I'm not a lawyer but I'm pretty sure this was a trade mark issue. He just had to avoid marketing his product using someone else's name.

In that particular case if you have an enterprise licence, Adobe have accepted responsibility:

>Adobe is so confident its Firefly generative AI won’t breach copyright that it’ll cover your legal bills The offer is available only to users of its enterprise Firefly product, which launches today

https://www.fastcompany.com/90906560/adobe-feels-so-confiden...

If they are that confident in their product why do they only indemnify users of their enterprise Firefly product?

Smells to me like a bad case of corporate marketing bullshit.

> Smells to me like a bad case of corporate marketing bullshit.

It is likely to be the large companies that worry about getting sued and tell their employees not to use the functionality. By offering protection Adobe get to sell more to the people paying millions a year.

I hope it happens. I'd love to see a market for selling training licenses to IP. This could be a small but real source of passive income for artists, authors, poets who don't mind that their IP being used in training sets. I wouldn't be practical to negotiate individually with each artist but I could see something work with larger collectives than can vouch for the quality of their members. Think like publishers, galleries, guilds or unions. It could offer a license and then share the proceeds with all members.

It's just flat out unethnical for LLMs to just soak up all this data, even off torrent sites(!!!), without any consent or agreement with the IP holders. Some model like this could be a win for everyone

> I think this will be a bigger issue than some people think. Maybe there's a market for 'clean' training data that doesn't include potential copyright claims.

Until this issue is resolved, that will have some value as risk mitigation.

Once it is resolved, it will either be a complete non-issue or an issue related to a much more knowable cost/benefit tradeoff.

> We'll know it's an AI because it talks like a late 18th century/early 19th century writer?

A mix of that and US government publications (which are categorically not subject to copyright).

there's a market for 'clean' jurisdictions that don't consider training neural networks to violate copyright, and japan has already declared itself such a jurisdiction

I think you mean 19th/20th, but that would be quite hilarious.

Dear Mr. Smith:

Your employer has generously agreed to offer you a position, quite reasonable, and with many great benefits at the venerable firm of 'Zumba'. Our interest is that you should join our staff forthwith and at the earliest date. A cab and man has been sent to retrieve you and bring you to our offices to sign all the necessary documents. Our offer is for a monthly stipend of five pounds, two shillings, and sixpence to be paid at the end of the month.

Thank you,

Most Humbly,

Hirebot 2347

I think the onus should be on Sarah Silverman to prove which inputs were hers and which outputs leveraged those inputs.

I think she should pay all of the court costs if she fails to do so.

Well yeah - that's how the legal system works assuming it gets all the way to court. In reality, Meta and OpenAIs lawyers will do a risk evaluation against the strength of the claim and if there's any merit at all there will be a quiet settlement.

IANAL, but my understanding is that in the United States, a civil defendant who prevails is not usually entitled to have their legal fees paid by the plaintiff.

[0] https://porterlaw.com/obtaining-attorney-fees-in-litigation-...

how could someone "prove" which inputs and outputs of a large ML model leveraged any specific data?

That's her and her legal team's homework assignment.

no, it isn't

it's an unsatisfiable requirement, and unnecessary to substantiate the legal claims

it's dumb to talk about

> it's an unsatisfiable requirement,

There's a wealth of primary literature describing means to probe models for the training data. There's also discovery and a whole host of other processes to answer this question.

The plaintiff that filed the lawsuit should prove what they allege.

> unnecessary to substantiate the legal claims

Why?

What if the model has absolutely zero of her data in it? Should she even be allowed to bring this case to court?

> it's dumb to talk about

Absolutely not! It's central to the entire case.

Even if her data is in the model, there's still a question of whether or not she should be compensated. I'd argue no for the same reason that babies that grow up watching Disney don't owe their entire intellectual output to the company.

Reply on news.ycombinator.com