Is AI Theft?

A Shallow Dive into the Ethics of Training AI Models

May 14, 2024

AI Stealing Art. Ironically created by DALL-E.

I was a little taken aback by some of the feedback I got after my “Adventures in Prompt Engineering” article. It was intended as a fun post illustrating the frustrations of working with new technology that won’t cooperate—something we can all relate to. However, some readers were upset, commenting that AI was “profiting from the theft of real artists...”

I’ve certainly seen articles claiming AI steals from artists,1 but I initially dismissed this as typical media scare tactics that pop up with any new tech. In the case of Andersen v. Stability AI Ltd, the lawsuit against AI image generation companies was dismissed, with the judge ruling there was “no substantial similarity” between the artists' original works and the AI's outputs. However, I’m starting to wonder—is AI actually stealing? What does it even mean to steal something?

Case 1: Stealing Physical Property

The most obvious example of theft is when someone takes something physical—your wallet, your car, etc. If I steal your TV, I have it now, and you don’t. Clearly, that stinks for you because you’re out of a TV.

Case 2: Stealing Intellectual Property

Then there’s stealing intellectual property. Copyright theft, patent infringement, and plagiarism all hurt the victim. In the workplace, someone swiping credit for your idea might get promoted ahead of you. Even among a group of friends, stealing someone’s great idea can boost the thief’s social status at the victim’s expense.

Case 3: Stealing Digital Goods

Stealing digital goods is trickier. For an up-and-coming artist, illegally downloading copies of their work hurts their ability to make a living. The same is true for established artists, but past a certain point, there’s diminishing returns. Take Taylor Swift. If a person copies one of her songs, she’s unlikely to experience financial hardship or lose her reputation. However, that wouldn’t be the case if everyone started doing it. This is why we have laws to prevent that kind of theft.

Things get morally grayer when we expand our definition of digital goods. In “Life After Capitalism,” I discussed how we created artificial scarcity to prop up capitalism in the face of digital goods that could be freely copied. Imagine if we developed technology akin to Drexler’s nano replicators, capable of copying any physical object at nearly no cost. Would we still justify artificial scarcity then? What if you could copy food as easily as downloading a Taylor Swift song? If this technology allowed us to copy food at minimal cost and distribute it to those in need, we could end world hunger. However, the manufacturer of the original food wouldn't be compensated, even though their actual stock wouldn’t be depleted—it's just a copy, after all. Would this act of copying food, without depleting the original, still be considered theft, just as with illegally copied music?

The argument that this constitutes theft hinges on the potential lost revenue, which a company could have used to pay workers or fund other beneficial projects. However, that’s old timey thinking, because in this scenario, the company doesn’t need money to pay workers. AI and replicators handle all production and labor. With artificial general intelligence (AGI) and advanced replicators, we enter a world of abundance, where no one needs to work and goods are plentiful. In such a world, artificially restricting access to goods to maintain scarcity is not just outdated—it's outright cruel.

Case 4: Training Artists

Growing up in the Whaling City of New Bedford, Massachusetts, I was surrounded by maritime art. Paintings and scrimshaw depicting whaling vessels out to sea and sailors armed with harpoons were a common sight throughout the city. In middle school art classes, I often found myself attempting (unsuccessfully) to replicate some of these iconic works.

That’s typically how young artists start their journey—they begin by copying the works of masters. Vincent van Gogh honed his skills by replicating paintings by Jean-François Millet. As a child, Wolfgang Amadeus Mozart extensively studied the works of Bach and Handel. Hunter S. Thompson once retyped F. Scott Fitzgerald’s “The Great Gatsby” and Ernest Hemingway’s “A Farewell to Arms” to get a feel for the rhythm and flow of their writing. Comedian Joan Rivers openly acknowledged her stylistic influence from Lenny Bruce.

Generally, copying the style of a master isn’t considered theft, particularly during training. No one views Van Gogh, Mozart, Thompson, or Rivers as less brilliant because they honed their skills by imitating the works of masters.

Case 5: Training AI

How does AI learn to write a novel, paint a sunset, or compose music? In many ways, it mirrors the approach taken by human artists. AI examines examples of what we consider good artwork and attempts to replicate the patterns it sees. Whereas an artist's neurons fire to help them recognize and replicate successful patterns, AI employs artificial neural networks for the same purpose. Just as an artist develops their style through continuous practice and reflection, AI refines its models based on the effectiveness of each output, progressively learning to not only mimic but sometimes even innovate on human-like creativity.

It’s important to note that AI doesn’t work like a camera or photocopier. It’s not actually copying the artwork, but rather generating entirely new art using patterns its seen before. In a nutshell, it’s using math and statistics to try and guess what patterns will lead to the best outcome. For example, in large language models it uses statistics to predict the next word in a sentence. If you start with the words, “I like eating…” and then ask it to continue the text, it will consider its options:

I like eating ice cream
I like eating broccoli
…
I like eating automobile tires.

It’s seen billions of examples of text, so it knows that you’re unlikely to say “I like eating automobile tires.” Statistically, “ice cream” is more likely the next text rather than “broccoli,” which is in turn much more likely than “automobile tires.”

I got into a disagreement with a lawyer once about whether AI was copying things. He was under the impression that ChatGPT was actively searching the internet and pulling text directly from sources, which would indeed be a clear violation of copyright laws. He believed this because ChatGPT could accurately produce text like song lyrics and the Pledge of Allegiance. For example, if you ask ChatGPT for the lyrics to the Backstreet Boys song Everybody, you get the following response.

For those who aren’t familiar, that’s an exact copy of the Everybody lyrics. Naively, one might think my lawyer friend had a point. However, ChatGPT does not search the internet when generating responses. Instead, it draws upon patterns and data it learned during its training phase, which involved a large dataset of diverse texts from which it learned how language is typically used. It generates responses based on this training, not by accessing or copying internet content in real time.

Asking for popular song lyrics is not the best test for whether AI is copying text directly from the internet. There are millions of websites that ChatGPT has been trained on, many of which contain the lyrics to Everybody, since it was a hit song. If ChatGPT didn’t have access to the internet, it would still likely remember the lyrics because it had seen it numerous times while being trained. Statistically, it knows that the word after “Everybody” is almost always “Yeah!”

A better experimental test is to see if it can find the lyrics of a lesser known song, one that is still easily searchable but unlikely to have a large presence on the Internet.2 For example, if we instead ask ChatGPT for the lyrics to the Backstreet Boys less popular B-side song That’s the Way I Like It, we get this:

This one start out close, but ChatGPT quickly starts making up lyrics after the after the third line. Even when explicitly instructed to copy text, ChatGPT has no choice but to come up with new lyrics! To be clear, this type of hallucination is a problem of its own, but it clearly shows that the AI is not simply copying text verbatim.

From this example, it’s clear that training AI is not fundamentally different from training a human artist. Human artists study existing works and train themselves by observing and mimicking patterns. AI does the same. It analyzes vast amounts of artistic data to learn how to generate its own creations. If we don't label Mozart and Joan Rivers as thieving plagiarists, then we shouldn’t with AI either.

Does This Mean AI Is Good?

No. Just because AI isn’t theft, doesn’t mean it’s all good. The major difference between AI artists and human artists is that AI can do things much faster and cheaper. There may be a distant future where we have nano replicators that can copy food and feed artists as no cost, but we’re nowhere close to that today. We’re entering a period where AI excels at specific tasks, putting certain jobs, like artists, at greater risk. This is one of the reasons why AI was a big point of contention in the 2023 Writers Guild of America strike.

The truth is, AI can already copy not only an artist’s style but also their voice and likeness. In 2023, music fans were abuzz with excitement over a collaborative track between Drake and The Weeknd, only to discover that the song in question was AI generated with no involvement from the artists. Another intriguing trend on YouTube is AI “resurrecting” deceased musicians. You can find AI-generated performances like AI Freddie Mercury singing I Will Always Love You, AI Whitney Houston taking on Thriller, or this gem:

Someday, we may have the world that futurists envision—a world of abundance and leisure enabled by AI and nano replicators. In this future, people can spend their time creating art and music without the fear of going broke doing so. However, until that day arrives, AI has the potential to create significant inequities. Even if training AI isn’t technically theft, the technology could disproportionately benefit those who control it, at the expense of those who contribute to its learning. We need to find ways to ensure that the technology benefits all of society fairly, safeguarding against the amplification of existing disparities and fostering an environment where innovation contributes to the common good.

So, what do you think? Is AI stealing from artists? Share your thoughts in the comments below. Also, if you enjoyed this discussion, don't forget: It costs nothing to subscribe for more insights and updates!

Diary of Numbers and Nanotech

Discussion about this post