WILL THE LAW STOP GENERATIVE AI?

Written, directed and narrated by Kirby Ferguson. Video commission for Bloomberg Law.

AI has historically been a bit of a joke.

It's been around since the fifties and despite all its big promises of human level intelligence, it often seemed ridiculous. Until now.

Now AI is a runaway train.

And the engine powering this surge of innovation is generative AI, a form of machine learning that creates original text, images, and code based on plain english prompts from the user. Generative AI is pumping out the hits.

• ChatGPT is an unprecedented sensation and large language model technology is being integrated into everything from search engines to the educational site Khan Academy.

• Image generation software like Midjourney and Stable Diffusion is producing amazing results that surpass human artists in some ways.

• AI music is rapidly maturing. An AI generated track featuring imitations of two of music's biggest stars, Drake and The Weeknd, went viral. The results might not be good but they're not exactly bad either.

AI is the hub of growth and excitement in tech. Forget about crypto, NFTs and the Metaverse – what was that anyway? The future is AI.

But this technology's capacity for harm is already apparent. AI can spew out fakery and nonsense, it might put a bunch of us out of work, and it arguably rips off human creators.

What, if anything, can stop AI or at least slow it down?

The answer is... the law. In particular, a rather unlikely branch of the law: copyright.

TITLE CARD: WILL THE LAW STOP AI?

The legal landscape of generative AI is mostly uncharted. There are major controversies in two areas: input, the human-created data that the AIs train on, and output, the text, images and code that the AIs create. Here’s Bloomberg Law IP reporter Riddhi Setty. 

Riddhi
When it comes to the input, artists are confused and worried that their art, images and text or other things are being used without their permission. On the output side, people aren't really sure whether or not they can copyright or register the works that they've created using AI.

The output side is a bit less mysterious because the copyright office has basically stated that copyright is limited to... humans.

Back in 2014, they declared that this monkey selfie cannot be owned by the monkey. Sorry, no royalties for you, little buddy.

And more recently, Kris Kashtanova was unable to copyright the images in their comic book because they were AI generated. But this is still a bit murky because the office also stated tjht an output could be copyrighted if there is sufficient human authorship – whatever that means.

They also said the prompts might be copyrightable. And yes, you can currently buy AI text prompts.

But these controversies are relatively minor compared to the battlefield that is the input side of image generation. Basically, AIs ate the internet, analyzed it and can now spit out an infinite array of new works that are some kind of transformed synthesis of human work. This process could possibly be illegal. To settle this grudge match, let’s square off two heavyweights. I mean that figuratively. 

In this corner, Mark Lemley. He thinks training AIs is not copyright infringement and he argues that AIs are really just doing something us humans do.

Lemley
If I want to write a fantasy novel, I can't write a Harry Potter book, but I can read a Harry Potter book and hundreds of other books and get a sense of what's in the genre, what works for literature, and what doesn't. We're giving machines the same rights that we already give people.

And in this corner, Ian B. Crosby. He thinks the way generative AI models train is copyright infringement. And he thinks what the AIs do actually isn't anything like human learning.

Crosby
What they're really doing is creating a collage out of little teeny bits of the original expression, and that the models themselves don't have within them anything like a general idea of the underlying facts.

There’s no dispute that AI’s training on copyrighted images is some sort of legal issue.

Crosby
The legal problem with generative AI is that in order to train a generative AI, you need to copy many, many, many works.

But Lemley adds that if you avoid copyrighted work altogether, and limit training to works in the public domain, you could end up with AI that sounds like a very old movie. 

Lemley
You could have ChatGPT, I guess, train on only text written before 1928. Anything written before 1928 is out of copyright. But it's not going to give you the same story. It's going to give you a very different sounding answer to any question you give. 

Crosby believes some of the most popular AI models wouldn’t exist as we know them, without infringing copyright protected work.

Crosby believes current generative AI cannot train, and thus cannot exist, without infringing the copyright protected work of human creators…

Crosby
Going forward, for these models to improve, you're either going to need better models, better training methods with the existing data that's out there, or you need more human creativity.

But in the past, courts have allowed new technologies to infringe the work of human creators, if that new technology also has a substantial noninfringing use, Lemley explains.

Lemley
Every time in the past we've had a new technology that has messed with those expectations in the media space, the existing players have said, "Oh my God, this has to be shut down," All of those things disrupted the market, but they also brought a lot of new opportunities to the world. I think we're going to find the same is true of generative AI.

One of the highest stakes cases now underway is Getty's case against Stability AI, which used Getty stock images to train its image generation software, Stable Diffusion. Crosby, remember he favors the people making the content AIs are trained on, argues Getty’s case is strong because the AI generated images are competing with Getty’s original images. 

Crosby
In the case that we have here with Getty Images suit, for example, where you're training the AI on stock photographs, which have a licensing market, and the output of the model potentially competes with those stock photographs. In other words, instead of going to Getty and searching for an image description and licensing that image, you can type a description into the AI, and it will produce a photorealistic image, which you can use in place. That's really problematic.

But Lemley argues that the purpose of training AI is not the same purpose that stock images serve.

Lemley
The AI is not interested in it for the purpose Getty uses it for. They are interested only in learning from it. The purpose is entirely different.

What's at stake in upcoming decisions like this is potentially the very future of AI.

Lemley
If we are unwilling to allow generative AI to train, and we say, "No, go get permission for all of the 2 billion images you use or don't run this at all," then things just shut down.

Ian Crosby disagrees. He thinks generative AI models could train on licensed and public domain works.

Crosby
In fact, there is a market that is emerging for this. And getting Getty, for example, others, Adobe, there are stock photo and journalism companies and newspapers that are starting to actually license their bodies of work. And I think OpenAI has actually been going out there trying to get some of these licenses.

Lemley warns that excessive restrictions on new technologies could hamper American innovation.

Lemley
I think there's a real risk, as there is with any new technology, that if we shut it down, if we say, "No, this is not permissible," before we realize all the things it's capable of, we never actually learn what we're missing.

While Crosby cautions that AIs are dependent on human creativity, and if AIs drive human creators out of business, that’s the worst scenario of all.

Crosby
If you train another AI on the output of that AI, and then you train another AI on that output, it quickly descends into gibberish. Human creativity is the necessary fuel for training these models. And if these models essentially wind up putting human creators out of business, they really are going to wind up eating their seed corn.

Congress is exploring AI regulations and there are several lawsuits now pending. But it may be a while before we get legal resolution. Whatever these decisions from the courts and congress are, they will define the future AI landscape.