Generative AI's Output: How Is It Created, and What IP Rights Should It Receive?

Generative artificial intelligence (AI) first performs broad, deep training on human-created materials, then, when given a prompt, can produce "original" works as output. Examples include ChatGPT (text), Stable Diffusion (images) and Video Diffusion (video).

Various disputes have arisen over generative AI. Sarah Silverman and others have brought a suit against OpenAI (producer of ChatGPT) for allegedly training its model on their creative works, and Getty Images has brought a suit against Stability AI for allegedly using the company's images to train a Stable Diffusion AI model. These suits allege copying at the AI's first step: absorbing human-created creative materials as input during its model training. Put simply, these suits are focused on the AI's alleged copying of materials as input.

This post focuses on the AI's output. That examination is much more complex because it requires a dive into how these models are trained and operate to produce their output. It also forces us to examine the mysteries of human creativity and how humans take inspiration from their own input to produce our creative output.

The "Black Box" of Model-Based Generative AI

ChatGPT is an example of a large language model (LLM), which we covered in an earlier post. The simple way to explain what the AI is doing is to say that it is predicting the next likely words or phrases: By reviewing the words that came before it, the AI picks the next word or phrase and the next, and so on until an answer is complete. Because the AI model was trained on materials that we regularly read (e.g., Wikipedia, e-books), these answers have the air of truth to us.

The reality of how that happens is, of course, much more complex. LLMs build massive networks of weighed vector nodes that focus on how words relate to others. The number of nodes in an LLM can be on the order of tens of billions, and these nodes form a vast neutral network of connections that represents the AI's "model."

When software uses that AI model to generate responses, the process is an unusual one because the model is a "black box" to engineers and researchers. This is in contrast to how software usually operates. Software is typically a perfectly "knowable" system because it executes a series of specific commands that humans tell it to – it is a walled garden of understandable functionality. Software engineers debugging an application, for example, can watch in real time as data passes through a series of commands while tracking changes to variables and observing which paths the software takes. There is a perfect audit trail of how the software moved from its input to output.

Generative AI models obstruct that view into how software moves from A to B. As a result, AI engineers often cannot explain why or how the AI generated its output. As one researcher put it: "If we open up ChatGPT or a system like it and look inside, you just see millions of numbers flipping around a few hundred times a second . . . [a]nd we just have no idea what any of it means." There is no usable audit trail.

The "Black Box" of the Human Mind

While ambiguity regarding how a computer reaches a result is unusual, it common in our everyday lives. At any given moment, we are making choices and generating output without the ability to accurately trace how we arrived at that output.

As an example, we can ask ourselves why we bought our last car. Because cars are expensive, we tend to spend more time considering those purchases than others. You might have weighed a mix of several factors, including:

1. It was within your price range.
2. It was the right size.
3. You have a favorable opinion of the brand.
4. You liked the color.
5. The mileage was acceptable.
6. A friend has the same car and likes it.

We might be able to build this into a prioritized list, but could we assign precise weights to each factor? More broadly, could we evaluate the influence a car commercial you saw earlier that year had on your choice? What about the fact that our aunt had that brand of car decades ago and thought it was reliable? How about when a car of the same color cut us off on the way to work? To what specific extent did these experiences factor into our decision? It is impossible to ascertain a definitive answer because there is no reliable audit trail of how we made that choice.

As with AI models, our minds are complicated, so the exercise of determining why we do what we do is likewise difficult. We are, effectively, operating with our own black box problem. (If the comparison between generative AI and the human mind seem far-fetched, scientists have found similarities between how ChatGPT and the human mind operate, at least with regard to next-word processing.)

George Harrison encountered this "black box problem" when one of his songs ("My Sweet Lord") had a striking similarity to the 1962 Chiffons hit "He's So Fine." Harrison faced a lawsuit where he claimed that while he did not knowingly copy "He's So Fine," he did admit to having heard the song before. The judge credited Harrison's testimony and had this to say about how the songwriter's input may have resulted in this similar output:

What happened? I conclude that [Harrison], in seeking musical materials to clothe his thoughts, was working with various possibilities. As he tried this possibility and that, there came to the surface of his mind a particular combination that pleased him as being one he felt would be appealing to a prospective listener; in other words, that this combination of sounds would work. Why? Because his subconscious knew it already had worked in a song his conscious mind did not remember. Bright Tunes Music Corp. v. Harrisongs Music, Ltd., 420 F. Supp. 177, 180 (S.D.N.Y. 1976).

In other words, the judge's ruling acknowledged that the human mind is complex and performs unknowable processing of input to produce output. Based on the low likelihood that two authors could produce such similar songs, subliminal copying seemed to be the most likely explanation of what happened, but no amount of discovery could prove it to a certainty.

If AI Is Operating Like a Human Mind, Should That Shape IP Policy?

This is another exciting time for intellectual property (IP), where a new technology arises to challenge how IP law operates. For the first time, we have output from AI that is creative and original. And that AI is producing its output in complex, unknowable ways. At least as of now, the fact that some generative AI is operating like a human mind is having no effect on IP policy. This is primarily because AI is not human and, as such, is incapable of receiving IP rights.

For example, generative AI has called into question the copyright ownership of its creative output. At present, however, the U.S. Copyright Office will not consider AI as an author because it is not human. Because there is no author, the raw output of generative AI is instead dedicated to the public domain. (The Copyright Office is now seeking public comment on these issues.)

As another example, parties have attempted to anoint AI as patent inventors but have so far found no success for the same reason. In patent prosecution, too, an applicant must be human.

If the gating issue is whether the AI is "human," what happens over time as the AI's processing becomes more and more humanlike? We are already seeing AI mimic the functions of human thought in small (but remarkable) ways. It is unlikely that a legislative body or court will ever bring AI within the definition of "human" It is possible, however, that AI will receive some specialized rights in the future. For example, corporations are clearly not human, but they enjoy many of the same rights as citizens because allowing as much legally eases their operation. As the law struggles with how to handle the emergence of AI, IP may be the first field where AI receives differing treatment than does traditional software.