Using code, photographs, documents, or other material to train a model isn't copyright infringement. The person operating the model is not violating the exclusive rights of the copyright author: they are not making copies or derivative works.
Any other result means that all AI development based on training models is going to grind to a screeching halt, because essentially all training material—text, pictures, recordings—is copyrighted.
How do they not make copies? Do you know how a computer works? Ever heard of RAM? (At least the German Urheberrecht recognizes this clearly: You can't do any processing on any data with the help of a computer without at least making temporary local copies, so there are exceptions to some rules. I'm quite sure common law copyright also recognizes this!)
Also the claim that this is not a derivative work is actually one of the disputed claims here…
> Any other result means that all AI development based on training models is going to grind to a screeching halt, because essentially all training material—text, pictures, recordings—is copyrighted.
Exactly, it's all copyrighted! That's why you can't use it for whatever you like. That's the whole point of copyright.
As a result this means that whoever wants to exploit that work in said way needs to buy (or get otherwise) a license!
Nobody said that feeding AI with properly licensed work would be problematic. Only the original creators need to get their fair cut form the outcome of such process.
I think that losing this lawsuit has much more serious consequences for Copilot than just having to connect to a list of millions of potential copyright owners - it would mean the model behind it is essentially a failure.
Personal opinion: the real situation lies somewhere in the middle. From what I’ve seen, I think Copilot has some ability to actually generate code, or at least adapt and connect unrelated code pieces it remembers to respond to prompts - but I also believe it just “remembers” (i.e., has a close-to-lossless encoding of the input) how to do some operations and spits them out as part of the response to some prompts.
I hardly think the lawsuit will really explore this discussion, but it sounds like a great investigation into what DL models like transformers actually learn. For all I know, it might even give insight into how we learn. I have no reason to believe that humans don’t use the same strategy of memorising some operations and learning how to adjust them “at the edges” to combine them.
In the end it's just a machine. It's not a person. So trying to anthropomorphize this case makes no sense from the get go.
Looking at it this way (and I guess this is the right way to look at it from the law standpoint) Copilot is just a fancy database.
It's a database full of copyrighted work…
How this database (and it's query system) works from the technical viewpoint isn't relevant. It just makes no difference as by law machines aren't people. End of story.
But should the curt (to my extreme surprise) rule that what MS did was "fair use" than the flood gates of "fairuseify through ML"[1] would be open. Given the history of copyright and/or other IP laws in the US this just won't happen! The US won't ever accept that someone would be allowed to grab all Mikey Mouse movies put them into some AI and start to create new Mikey Mouse movies. That's the unthinkable. Just imagine what this would mean. You could "launder" any copyrighted work just by uploading and re-querying it form some "ML-based database system". That would be the end of copyright. This just won't happen. MS is going lose this trail. There is no other option.
The only real question is how severe their loose will be. They used for sure also AGPLv4 code for training. Thinking this through to the end with all consequences would mean that large chunks of MS's infrastructure, and all supporting code, which means more or less all of Azure, which means more or less all of MS's software, would need to be offered in (usable!) source to all users of Copilot. I think this won't happen. I expect the court to find a way to weasel out of this consequence.
[1] https://web.archive.org/web/20220121020414/fairuseify.ml/
With current technology, the only licensing model we can offer is "give us your training set example, we'll chuck a few pennies at you out of credit sales and nothing more". We can't even comply with CC-BY because the model can't determine who to attribute to.
If an AI model is the joint property of all the people who contributed IP to it, it’s a pretty hugely democratic and decentralizing force. It also will incentivise a huge amount of innovation on better, richer data sources for AI.
If an AI model isn’t joint property of the IP it learned then it’s a great way to build extractive business models because the raw resource is mostly free. This will incentivise larger, more centralised entities.
Much of the most interesting data comes from everyday people. A class action precedent is probably good for society and good for innovation (particularly pushing innovation on the edge/data collection side)