AI, Art & Intellectual Property with Kent Keirsey, Founder & CEO at Invoke

So maybe how should AI practitioners and leaders be thinking about intellectual property when it comes to AI? Yeah, I mean, I think when we're talking about intellectual property, it typically comes back to the training of the model as well as the outputs that come out of these models. The models themselves are in some sense a new form of intellectual property. So we're kind of having conversations about licensing on the weights, what you can do with the weights, how you can fine tune them and train them. But I'll kind of break these down into a couple of different pieces. We're talking about the inputs that go into training these models. There's a lot of controversy right now around whether utilizing data that is pulled from the Internet and the Wide Web and then using that to train a model, whether that is considered fair use. Most of the intellectual property law that this falls under is really the kind of the copyright law. And I've had conversations with a lot of experts in the field. In fact, we just recorded a video with Matt Sag, who recently, in July of last year, sat on the Senate hearing committee around this topic. But the way that he frames it, and I think this is a really, really interesting way of thinking about it, is that a lot of priority cases have been settled with this notion of non expressive use. And really what that means is when you're using the data and creating a copy for training, you're not using the expressive qualities of that image, you're not copying it to produce something that is replicating the content. You're not trying to create an overfit model, you're utilizing it in a non expressive way. And that typically does fall under fair use. And fair use would allow us to effectively train models on anything so long as they are publicly made available. We're not breaking the law with access to the data. We're not like taking pirated copies of books, for example, and training models on those. This is just like general web data. And I think that's the big area that we think is matters right now is focusing on how models are trained and getting some coverage there, because we don't want to train models that are created illegally. But I think most of the legal professionals are moving towards this understanding of this as being a fair use activity. And I think the fact that we've seen large companies, even like Apple, create models that have been trained on data that's been scraped effectively means that the large majority of the American corporate populace, if you will, they've made the decision that this is likely the way that it's going to go. And I think there's a whole host of geopolitical reasons that it has to go that way. Those are, those are the areas that we focus on when we're talking about inputs, when we're talking about outputs, and how you get kind of protection on the outputs. But both the model as well as the things that those models generate, we're getting into an entirely different problem space. And I think with models you have to focus on the license and what you can do with that. We recently saw with stable diffusion 3, for example, in the image generation space, the license itself will actually cause most enterprises and corporations to back away if it's not a friendly license. Openly licensed, permissive, allows for fine tuning, training, owning the asset that is key for open models with the outputs of generative models. We're asking questions right now like how do we get copyright protection on the things that we create with AI? What can we protect from a patent perspective if we co create something? And the Copyright Office right now has basically been very conservative with what they're going to grant copyrights over. Their argument is that at least the early technologies in this space lacked sufficient human expression to merit a copyright. So if you go in to Dall? E and you type show me a picture of a dog, that dog is not copyrightable by you because the amount of expression that you could put in by show me a picture of a dog is minimal. Right? And that's where I think today we're talking about control and privacy as big focuses. How do I control these models such that I can get some amount of human expression there in the process and then claim that expression is copyrightable? And then from a privacy perspective, when we create assets like let's say a model, like a fine tune, how are we making sure that we are the ones who benefit from that and not the AI companies? Are they training on our data? Are they training on our ip? When we put it into their system, is it making their model better or is it making our model better? Are we kind of the owner of our own destiny?