Hacker News Reader

Comment by jprete | Hacker News Reader

The article weasels out at the end by claiming that companies “may be unable to comply” with requirements to delete personal data. It’s easy to comply - if there’s no other choice then you delete the model and all backups and derivative data that was trained in flagrant violation of the law.

View on HN

For most companies “deleting the model” is equivalent to dissolving the company so that is equivalent to not being able to comply. More realistically what they would need to do is exit the market of the country that has such stupid laws.

>For most companies “deleting the model” is equivalent to dissolving the company

I'm okay with that. It's sad that you're not. If you're willing to start a business on such shady foundations, there's a really good chance your business will continue to make shady decisions in the future. It's better to find and remove the cancer early

I just disagree that it's shady. If you put your personal info into public circulation I think it's absolutely morally fine to have a model train on it.

There's a difference between being in public and being available for public use. Books in a book shop are "in public", but only those out of copyright are available for public use. Similarly, blog posts are in public, but not being given away for anyone to resell or include in their product.

For an image/photography example, I may have my photo taken at an event and sign a release or agree to the event's T&Cs that say they can put publicity photos on their social media or website, but I don't intend that my photo will be used by an unrelated party to generate images of my face (this may not actually be restricted, yet, but I'd prefer it to be and I suspect few would consider it a reasonable outcome).

> There's a difference between being in public and being available for public use. Books in a book shop are "in public", but only those out of copyright are available for public use

Copying books can be fair use if it is sufficiently transformative, copyright doesn't block all use of copyrighted material. A major case on this was around Google building a search index for books:

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

Open AI is currently attempting to argue that training a model is more transformative than building a search index, whether they win that argument or not is still to be seen.

This is true for some types of transformative work, but aggregating to make money off others work is not generally allowed. I think Open AI are in for a hard time with this when people look at the results and see their work clearly.

Building and granting access to a generative model is far more damaging to the market prospects of the infringed works, which is also a factor. Like…the whole goal is for it to be able to replace what humans do, which is about as market-damaging as I can imagine. I’m not a lawyer, though.

your data can be public without your consent, what then ?

I may sue the person who made it public without my consent.

Which law is that?

The post that you're responding to doesn't make any claim about the law; it expresses that the defense that an AI company might be "unable to comply" with a command to remove user data is not an honest one.

"flagrant violation of the law" heavily implies that OP has some strong feelings around this, and I want to know what law they think is potentially being flagrantly violated.

There is currently no law in the US or EU that I'm aware of that consider models trained on PII or copyright data to be considered as including that data. This is the core issue everyone is discussing about copyright around models and whether publicly scraped data is fair use for these purposes.

Until this is decided, likely by a long series of trials in high courts, or passage of law through the appropriate legislative bodies (and then likely after many challenges to it) will GDPR and CCPA be able to apply to these language models.

So no this is not currently a violation of any law.

I was daydreaming about other possible ways to comply, but I don't know if they'd really work in practice. Ideas like:

- Supply copywritten keywords and approved responses to try and retrain the model.

- Train the copywritten content again with inverse weighting. I imagine this one wouldn't really work well and still keep the model performance up...

Reply on news.ycombinator.com