People on another forum says portraits are personal data and use of them is a breach of Art. 6 GDPR Lawfulness of processing. This creates a problem in most European countries. This is a breach of privacy laws, and not a copyright issue.[1] Not sure how to interpret the local copyright law on this. It can be read both as it is legal (even to just repurpose all kind of images no matter license) and as it is illegal to do it (it would be similar to sampling of previously published music). Seems like you are allowed to train a model, but you can't publish it… [1] https://gdpr-info.eu/art-6-gdpr/ On Sat, Jan 18, 2020 at 11:58 PM Ryan Merkley <rmerkley(a)wikimedia.org> wrote:

Thanks for that. Pardon me if I've missed something, but that seems to imply, but not directly state, that AI training is a derivative work; could you clarify that? On Jan 18, 2020, at 2:58 PM, Ryan Merkley <rmerkley(a)wikimedia.org> wrote:

This whole subject raises interesting and important legal and ethical issues, but are there any direct implications at this time for Wikipedia/Wikimedia projects? Newyorkbrad/IBM On Sunday, January 19, 2020, Ryan Merkley <rmerkley(a)wikimedia.org> wrote:

they work, I would say that your trained network is a derived work if the weights are learned from a specific training sample. If it does not learn from a specific sample it is not a derived work from that sample. It is not sufficient that a specific image is in a training batch, it must also trigger learning. The goal for the training is partial or full reconstruction of properties necessary for some operation. It is perhaps easier to see with an example: Imagine you have photos of Picasso's art, and you try to estimate whether some art is within his Blue Period. Is the network derived work from Picasso's art? And how does that compare to whether the network itself is a work of art, is it still derived from another (set of) work of art? I would say training of neural networks is a derived work built on the training samples, but I would not say itself is a derived work of art, even if it can be copyrighted. The most common solution these days seems to be to encode what is called "eigenfaces", and then use those to create a kind of hashes for similarity detection. Eigenfaces are a kind of arketypes of how a face look like, and mixing such faces creates new ones. It is a bit similar to those flipover albums the police uses, but allowing a lot more of gradual changes. Face detection is often visualized as vector points in movies, but this is not how eigenfaces work. Or rather, it creates vectors, but not as dots on a portrait. Note that ClearviewAI goes a good bit longer than just learn some variant of eigenfaces. On Sun, Jan 19, 2020 at 4:30 AM Benjamin Ikuta <benjaminikuta(a)gmail.com> wrote: