Artist finds private medical record photos in popular AI training data set

> "Late last week, a California-based AI artist who goes by the name Lapine discovered private medical record photos taken by her doctor in 2013 referenced in the LAION-5B image set"

> "...it bothers her that private medical images have been baked into a product without any form of consent or recourse to remove them."

> "Ultimately, Lapine understands how the chain of custody over her private images failed"

The article keeps claiming that private photos are included in a AI product, but the fact is that these were public photos when they entered the product. In fact, it's only because they were public that they were included.

The article attempts to frame this situation as "LAION steals private photos and won't take responsibility" but that's not only a lie, it ignores the real issue. If there is a problem here it is that these photos were made public in the first place when they shouldn't have been, at least according to the person photographed. It's entirely unclear what rights she has to these photos or to control how or where they are used, or that there was any illegal activity from any party in making those images freely available to the public.

Even if the LAION data set never existed it wouldn't change the fact that those images have been made public and have been being repeatedly copied/shared/viewed by all manner of people for any number of purposes. Calling out this one AI project is ridiculous. There are some good discussions to be had about if it's okay to use certain types of material in training datasets, but it doesn't help anyone to grossly misrepresent the issue.

I feel bad for people who suddenly realize that images of them are public when they didn't think that would happen, but there is notably zero information in this article about what exactly caused these images to get posted to public websites, who published them making the images public, and what rights the publishing party had when they did it.

Discussions on how to keep data that should be private from being made public, or making people aware that data they expect should be private may not be private at all is a lot more productive than hit pieces trying to help claw back images that have already been made public or spreading FUD about new and powerful technologies that only make use of already public data.

"publicly published" or "published and made freely available on the internet" perhaps since all sorts of data gets published privately with access tightly controlled where as having something exposed to the entire global internet is about as public as anything ever can be.

> If I surreptitiously take private photos of you, and then I publish them on the internet against your wishes, that does not suddenly make the photos any less private.

I still couldn't say those were 'my private photos', the photos themselves would be very public (and they'd be yours), although you could say that the subject would be private I suppose... If you broke into my house and set up the cameras I could say you photographed by private moments. So, yeah I guess that'd be considered publicly published/exhibited photos of private moments.

> I suspect you think logically, and you would like the words private and public to be disjoint sets, however in practice those words have overlapping meanings depending on semantics.

It gets even more semantically complicated if you broke into my house, found my photo album, stole my private photos, and uploaded scans of them to the internet. At that point I'd probably want to say you "stole my private photos" which could be true, although it might be more accurate to say "You stole what were my private photos and made them public"

That's the problem with the article. It makes it sound like the photos were private until they were added to the AI's training data.

A sentence like: "...it bothers her that private medical images have been baked into a product without any form of consent or recourse to remove them." might then be better stated as "...it bothers her that publicly published images that originated from private medical files have been baked into a product without any form of consent or recourse to remove them."

The distinction that the photos themselves were public matters since that explains how they ended up in the data set. I think everyone would agree that no one should use your private photos without your consent. Just gaining access to them would require something extraordinary making that easy to avoid.

I think fewer people would suggest that a web scraper/crawler should have to somehow discover the full backstory of every publicly available photo it finds online to determine if a photo depicts something someone might consider private. It's admittedly a flawed assumption that a photo published to the internet where anyone can access and download it is fair game, but I'm not sure what would work better. Pretty much every search engine works under that same assumption.