The Robot Authors

A little while back I wrote about the push to have artificial intelligence systems named as inventors in patent applications. This is going to get into a similar question in the field of copyrights: Can an AI system be named the “author” of a creative work?

If the idea of a machine coming up with an invention is challenging, the idea of a machine writing a novel or painting a picture may be easier to wrap your head around. The basic ideas of the Infinite Monkey Theorem have been around for over a century: If you let a monkey bash on a keyboard for an infinite amount of time, it will eventually produce the complete works of Shakespeare.

That works because “infinity” is so impossibly large that eventually you will get a randomly generated string that happens to match what Shakespeare wrote. By the same token, I can program a computer to randomly generate text and check periodically to see whether it happened to shakesp. Nobody looking at that from the outside would say that the monkey or the computer was being “creative.” On the other hand, we tend to ascribe creativity to human artists who use mostly random processes. Where is the line?

But I’m getting ahead of myself. This post is going to include the state of the art for machine learning in artistic endeavors, current events in the field, the requirements that copyright law imposes, and a punchline.

Super Good With Words

I briefly touched on the concepts of machine learning in the Robot Inventors post, but let’s dive a little deeper into how it works for text and images. In both cases, machine learning models have gotten seriously very good at generating new text and new images based on prompts. As Reginald from Nedroid would say:

In the world of natural language processing (aka, Talking To Computers), the cutting edge is in transformers, and specifically in the GPT-3 model. Boiled down, a transformer model takes text as a sequence of words and looks at the whole sequence, determining what parts are important and how they influence the meaning of other parts. I’m not going to pretend to get you up to speed on it all in this post, but let the takeaway be that these models do a great job of predicting what the next sentence should be given the last sentence.

One big place this comes out is in chat bots, which have long been a staple for artificial intelligence research. You say something to the bot, and the model computes the best response. With enough training data, you can get surprisingly human-like responses out of the machine. More on this below!

The other type of output I want to point you to is in image manipulation. Instead of working with representations of natural language, a machine learning model can work with images. They commonly use convolutional neural networks to process two-dimensional image data, which lets them recognize and manipulate patterns over an area of the image.

Recently efforts have been made to combine the two types of model, providing a way to translate between natural language inputs and image outputs. Basically, telling a computer in words what you want to see, and then the computer paints you a picture. Let’s look at some examples!

DALL-E, DALL-E 2, LaMDA, and Dr. Thaler

My goodness, there has been a lot of this in the news lately! Artificial Intelligence has been Big Meme for a couple of months now because multiple examples of its creepiness came out all at once. Let’s start with DALL-E and DALL-E 2, which have been the subject of a lot of social media posts.

The DALL-E models use GPT-3 to process text inputs and then generate images that match the input. So you could ask it to generate, say, “a man wrestling a walrus,” and it would do a bang-up job creating a photo-realistic image of exactly that. There is even a toy version available, which is not derived from DALL-E but instead attempts to reproduce it. The output is less impressive than the examples I’ve seen of DALL-E 2, but is still enough to raise some serious legal questions:

This is just the latest example of computers-making-weird-images. You may remember when the Deep Dream Generator was making the rounds a few years ago, and everyone was posting super-trippy selfies. It’s just that, now, the models are coming up with images from scratch.

Another big news item lately is in Google’s Language Model for Dialogue Applications (LAMDA). This is a chat bot that Google is developing using its own transformer models, and it has been the subject of A Fracas. The short version is that one researcher became convinced that the chat bot was sentient (in part because it told him so), and went to the news with that claim, and got himself suspended. And now everyone has Big Opinions on what it means to be sentient and whether this chat bot was sentient. The bot even wrote a story!

(As an aside, my personal opinion is that the things we think of as hallmarks of sentience just aren’t as special as we always thought. Also, humans are built to recognize patterns and look for meaning, so it is not surprising that some people will come away from an advanced chat bot feeling like they were talking to a human being. It is specifically designed to mimic us to create that effect, after all.)

The last bit of news didn’t catch headlines in the same way, but is super important for the theme of this article: Dr. Thaler is still at it! In my post about robot inventors, Thaler was filing lawsuits to establish that a machine should be listed as the inventor on a patent. Now, Thaler is suing to establish that his machine learning model should be considered the author of this image:

Thaler’s application for a copyright application, on behalf of the “Creativity Machine,” was rejected on the basis that it lacked an author, and that decision was confirmed by the Copyright Review Board in February. The Copyright Office’s decision gives a good overview of the issues at play, so here it is:

Thaler has now filed a lawsuit, appealing that decision.

A Modicum of Creativity

Let’s get into what’s required for something to be copyrightable. Here are the important bits of the statute:

(a) Copyright protection subsists, in accordance with this title, in original works of authorship fixed in any tangible medium of expression, now known or later developed, from which they can be perceived, reproduced, or otherwise communicated, either directly or with the aid of a machine or device. …

(b) In no case does copyright protection for an original work of authorship extend to any idea, procedure, process, system, method of operation, concept, principle, or discovery, regardless of the form in which it is described, explained, illustrated, or embodied in such work.

17 USC 102

Basically, it has to be an original work of authorship, and you can’t protect a bare idea. The Copyright Office requires human authorship, but the statutes themselves don’t define what an “author” is. As the Compendium of U.S. Copyright Office Practices, Third Edition states, “The copyright law only protects ‘the fruits of intellectual labor’ that ‘are founded in the creative powers of the mind.’” (quoting Trade-Mark Cases, 100 U.S. 82, 94 (1879)).

It turns out there is caselaw on the subject, though nothing that directly hits on a computer specifically as author. For example, a monkey can’t get a copyright, but because the monkey lacks standing to sue, not because the courts decided on whether a monkey could be an author.

Another key point of authorship is that they must include some “modicum of creativity.” In Feist Publications, Inc. v. Rural Telephone Service Company, Inc., the fight was over whether you could copy a phone book. See 449 U.S. 340 (1991). The Court held:

As mentioned, originality is not a stringent standard; it does not require that facts be presented in an innovative or surprising way. It is equally true, however, that the selection and arrangement of facts cannot be so mechanical or routine as to require no creativity whatsoever. The standard of originality is low, but it does exist.

Feist at 362

I personally see the “originality” requirement as not just a requirement that the output be something new, but also a requirement on the method by which the original work is created. Dr. Thaler naturally disagrees, and we’ll see where the courts land on the subject!

But also at issue are situations where a human being and a machine work together. Generally copyright law is not biased when it comes to the tools used to author a work. For example, we award copyright protection to photographers, even though they may be capturing a wholly natural scene. Courts have found that the photographers choices in how the scene is photographed suffice to establish authorship.

So when I tell the machine to make a man fight a walrus, do I deserve the copyright for that image? All I did was supply the bare idea, rendered in text, and machine did the rest. And as we know from the statute, an idea cannot be copyrighted, “regardless of the form in which it is described, explained, illustrated, or embodied in such work.”

In this specific case, I am tempted to say no. My creativity was involved in the uncopyrightable idea, not in generating the expression of that idea. But I did select the image above over eight less-good outputs, so perhaps my creativity came into play there? That might actually be a fun test case…

As all of this plays out, my hunch is that the courts will refuse to grant copyrights to purely AI-created works, but will grant copyrights to works generated by AI models in response to a human’s input.

[EDIT 10/12/2022 My position on this has developed in the months since I wrote this article, and I now think that my hunch on this point was not correct. I now draw an analogy to a client commissioning art from a freelance human artist. The client supplies general instructions to the artist, and the artist creates the art. In that case, the artist would hold the copyright and the client would not.

Using an AI model is similar: You supply an idea–even a relatively detailed idea–and the machine creates the art. You can sift through several outputs to pick the one you like best, but that’s no different from the client saying, “I don’t like this one. Do it again.” You are simply supplying the uncopyrightable idea, while the machine does the work, and so you do not possess the copyright. Even if you refine the prompt at each iteration, that seems no different from the client saying, “Do it again, but this time have the dog wear a bowtie.” You still aren’t supplying anything beyond the bare idea.

That said, if you perform some modicum of creative work on the AI’s output, I still believe that would be sufficient.]

The Punchline I Promised

Dr. Thaler is getting his time in court. His lawyer recently argued his patent case before the U.S. Court of Appeals for the Federal Circuit, and a decision is still forthcoming. I downloaded the oral arguments, which you can listen to here:

At the end, after Thaler’s lawyer gave the first argument and after the Copyright Office’s lawyer gave their response, Thaler’s lawyer had three minutes to rebut. Here’s the final exchange, where the court probed about copyright issues:

Judge: By the way, do you happen to know, do the mythic monkeys who type out Shakespeare get to be copyright holders?

Lawyer: Uh. Well. I see I’m past my time so if I…

All: <laughter>

Id. at 25:59

Someone is going to have to answer the monkey question eventually. And when they do, you bet I’ll be there reporting on the action!