In October, writer William D. Cohan published an opinion piece in the Washington Post titled “AI Is Learning from Stolen Intellectual Property. It Has to Stop.” First, the moment we think of ideas as “intellectual property,” we’ve accepted the conception of thought and its practical application held by property owners—and, of course, especially by those who own a lot of property. Second, the notion that what AI does is “theft” misunderstands how AI works, and for that matter, imagines for AI a wholly new standard regarding the taking in of information, a standard which criminalizes how AI works with information, even though human beings work with information (and use that information) quite similarly. Obviously, AI’s information processing is considerably faster and can handle much more volume than humans can…but if it’s not theft when a human being’s thoughts display the influence of another thinker, it can’t be theft when an AI’s output does so.
Humans can read things, look at things, listen to them, and take in information to understand it, to work with it, to combine it with their own ideas or others’ ideas, and so on. It makes no difference if the ideas they’re working with are “paid for” directly: if I read a copy of one of Cohan’s books that I checked out of the public library, even though the library paid for that copy, I am not stealing anything. Same if I borrow my neighbor’s copy of the book. Same if I’m glancing at the book at the bookstore, and something Cohan writes stirs up some thoughts of my own. I owe Cohan nothing if a stray phrase or sentence spurred on my own thoughts. If the influence is specific or quite direct, I may well cite him…but it may also be that the idea mixes in with all the other ideas I encounter and workplay with, such that in the end, it’s clear to no one what sources germinated any particular set of ideas.
“Permission” to take in the ideas of a book, a painting, a piece of music may, in some cases, be restricted…but when people borrow a book, see a copy of an art museum’s painting on a poster, overhear someone else listening to music, we don’t call that “theft.” We call it “learning,” or “thinking.” Every idea out there can hypothetically be traced back to some person’s original thought (or several people arriving at the idea independently), but there’s no intellectual accountant tracking every last thought and making sure its originators get paid, nor intellectual cops staking out information thieves thinking about and with ideas they did not generate entirely from ideas they somehow “paid” for.
More to the point: if Cohan’s work is online, and he has not specifically locked it down only to those who pay him or get his permission to read it, the fact that AI is “scraping” it to learn from it how language works is no more “theft” than my going to the library is. Yes, the library paid for the book…(and Cohan was, presumably, paid something to write his book, even if the book didn’t earn back his advance)—but no one else need do so.
And there are many situations in which we take in information without our paying a cent or having any specific permission. The situations in which another person’s thoughts must be cited (which still does not typically require either payment or permission) have to do with the use of specific words or phrasing, or particular ideas. A general mood, or feeling—the idea that, say, I want to write a song that sounds like a Chagall painting looks—has nothing whatsoever to do with “theft” or any kind of copyright law.
In describing the process of “scraping” information and analyzing its functioning, Cohan invents (as have many others recently) an intellectual property right from whole cloth. “Style” is not copyrightable. Furthermore, Cohan and others often write as if AI spits out discrete chunks of the texts it’s trained upon and digested. It does not. When asked for a text that’s similar to Dylan Thomas, it does not output a hash of individual Dylan Thomas lines or phrases. Instead, it analyzes the structure and nature of (in this case) Thomas’s language, and generates phrases that reproduce that structure and nature but are not phrases Thomas ever wrote.
That is, AI does not do collage. (And it’s an interesting point that collage artists historically are not expected to pay for, or even necessarily name and acknowledge, the source of their imagery.) It does not merely mix and match existing ingredients, as if putting together a stew from leftovers. Instead, it gleans certain traits or characteristics from its sources, which are used to further its ability to output language (etc.) that works similarly to those sources, but which does not quote or paraphrase or summarize those sources directly. (Note that Cohan did not find out his works had been scraped because he recognized some aspect of his work in any AI’s output. Instead, he was told that his work had been used. This is a huge distinction: if a writer plagiarizes paragraphs from a second writer’s work, it’s a form of theft…but if that second writer believes that the first writer was strongly influenced by his work, that is in no way actionable…and only the most crankish of writers would consider “influence” to be a form of theft.) Style is not copyrightable.
This is a key point. “Style” is a mysterious and intangible thing which, nevertheless, humans are fairly good at recognizing. “Fairly” good…but not at all unanimous. One person’s “influenced by Pixies” is another’s “what? I hear a lot of XTC but…Pixies?!?” (This is specifically about Bang! The Earth Is Round by the Sugarplastic, fwiw, and is based on an actual online conversation…)
Any artist is influenced by the books, paintings, music, and so on that they experience. Quite often they seek out work—or do research with such artwork—to enhance their knowledge and experience. It is not required that they pay a damned cent for this, nor acquire anyone’s permission. See: “public library.” Any artwork would be impossible without the artist’s influence and awareness of other artworks, both in the same field the artist works within and more broadly.
Furthermore: Cohan argues in bad faith by presuming one set of rules for humans, another set of rules for non-human information processing. Why? He accepts that people can and do take in ideas and information they do not pay for, and they do not need express permission from whoever might be expressing that idea to do so. And it seems he accepts this, since he never mentions it as being a problem…and he himself certain avails himself of such information all the time.
An irritating aspect of Cohan’s argument is the way he tries to manipulate readers into resentment by constantly referring to the extreme wealth and profit-seeking motives of the corporations funding AI research. Stirring up emotions is not good argument. It makes no sense to evaluate an action as bad only when a rich person or corporation does it if we have no problem with that action when a non-rich person or corporation does the same thing. Laws cannot be written that prohibit certain behaviors only if the legal person (individual or corporate) has wealth above a certain level.
Nevertheless: AI is an immediate threat to some workers. What should such workers do about such threats? What did Hollywood screenwriters and actors do? They went on strike. They made demands. They made arguments. But they largely focused on their skills and rights…not only alleged “thefts” of IP by AI. In doing so, their arguments against letting studios replace them with AIs was given depth and reason…rather than just flailing emotion.
It’s also true that certain commercial art, writing, etc.—the kind that exists primarily to be functional or make someone a lot of money—is fairly reproducible by AI, because AI does a good job of reproducing surfaces. And the point of such art is not that it be distinctively the product of a particular artist, but quite the opposite: that it be general and generic enough to seem like anyone’s and everyone’s sentiment and expression. Still, there are skilled professionals who create such functional art, and their careers are endangered by AI. It is important to protect the livelihood of such workers…but railing against AI as “theft” or otherwise pretending it can be made to go away are not among them. Such approaches are intellectually dishonest, ineffective…and in fact, accept key premises held by those who’d just as soon turn over their kitsch creation to AIs, human art workers be damned.
These “theft” and similar anti-AI arguments argue past the real issue, and do so in a way which accepts the terms of the ownership class in the first place. This is not a useful way to support artist workers. Consider what happened when abortion rights activists acceded to mainstream media descriptions of forced-birth advocates as “pro-life.” It’s hard to argue against a term like “pro-life”—who could oppose being in favor of life, after all? Even when it’s pointed out that only certain lives matter, that they’re dubiously “lives” in the first place, the rhetorical damage has been done. And the image of a group of people protecting life is set. Similarly: to accede to the notion that “art” is primarily a product of one’s labor, something produced while doing a particular job, accepts already that the artist is an employee. And that means that the work they’re doing is for someone else’s benefit, for someone who asserts a right over their time and work, just as is true of any other employee. “We need to get paid for our labor” is true…but the sheer existence of AI is not the problem here. The existence of labor exploitation is.
Not to mention that such arguments somehow imagine that AI can be eliminated, controlled, penalized, etc. Herding cats is hard—herding them back into the bag once they’re out of it is impossible.
Many of the proposed remedies to the sort of “theft” Cohan describes are likely to harm artists, such as enacting legislation that defines “style” as a copyrightable aspect of artwork. If such concepts were to be legislated, it would create a new property right in an area of creative work previously uncolonized by the concepts of property, ownership, and money: the nebulous notion of “style.” But the creation of such a new category of property would likely result in corporations rushing to copyright “styles” in advance of the actual human artists’ doing so. There is precedent—even in the absence of such rules, corporations have tried to monetize “style” and acted upon that belief. Consider: Geffen Records sued Neil Young in the ‘80s for not being a “Neil Young-like song generator.” And John Fogerty was sued—by the corporation that had acquired the rights to his music—alleging that he plagiarized one of his own songs (the rights to which now belonged to somebody else). Or consider the example of these YouTubers https://youtu.be/TsMMG0EQoyI?si=7TAkMxeUXAvJ07n2, in which a YouTube video using a musical composition clearly in the public domain was claimed in a copyright action, forcing the small artists involved to spend time and money defending their right to (in this case) put up videos of themselves playing Mozart. I’ve known musicians who’ve uploaded their own material to YouTube and been threatened with legal action by large corporations on the basis of bots who “think” the music is someone else’s.
The concept of “intellectual property” itself is a trap. The moment you accept that ideas can be treated like property, you put your ideas at risk of capture by richer, more powerful entities than yourself if they are interested in those ideas. The comparison of ideas to property is flawed in many ways…it would be better to think of ideas as “something humans do” which, when a particular human has an idea that’s distinctive, useful, entertaining, etc., requires recognition and, in some circumstances, a degree of compensation. Especially if such ideas arise upon demand of someone paying the human to come up with such ideas. However…the fact that the ideas came “upon demand” makes them more the specific person’s creation, not less. Because they are now the product of specific human labor…rather than just an idea a human came up with idly daydreaming one afternoon.
One thing that makes ideas unique is that, first, someone who gives them away still possesses them. Another thing that makes ideas unique is that they tend to be more functional, more interesting, more powerful, more entertaining, more moving, and so on, the more they are shared. And still more so when they are adapted, worked with, played with, fucked with, turned upside down, and otherwise used by other humans workplaying with ideas.
Scientific knowledge is maybe the best example of this phenomenon. Keeping a scientific discovery to oneself, thinking of it as “property,” is foolish and prevents the idea from attaining its value scientifically and in any other way. The same is true in the aesthetic realm: your poem, present only in your journal, has no value to anyone else. Sharing gives ideas value.
The other common overstatement, or fear, in arguments like Cohan’s is to extend what AI is likely to be able to do in an extremely unlikely manner. Non-commercial art, and/or art that moves beyond merely being functional, cannot be reproduced by AI. Such art incorporates substantial aspects of the human experience, which AI does not (and, I would argue, can never) have.
Human experience is based in that human beings are housed in a physical body, which experiences pleasure and pain, and which is finite and subject to decay and death. Such bodies are always located in space, in relation to other bodies and significant objects in the human’s world.
This is not true of AI, no matter how well it might ape “intelligence.” The situation is analogous to someone perfectly pronouncing every word of a speech in Italian…which they learned phonetically. Or even the same speech pronounced correctly and also bearing a fine, actorly emotional cadence…also learned from, say, recordings of the same speech. But the actor (in this case) does not understand what they’re saying (except insofar as they may have read a translation). They are merely offering a very good but rote reproduction. Human texts come with self-awareness and understanding. AI texts do not.
Even if AI did offer a mishmash of quotes (see above), it would still clearly not be a human. Even the more “creative” version we see that’s based on extensive analysis of how texts work (any text, whether verbal, visual, audible, etc.) still does not come along with self-awareness or any level of understanding.
But, hey: let’s imagine that, some day, AI genuinely achieves intelligence. Let’s say that such intelligence includes (as I believe it must) self-awareness, even will. Would not such AIs have rights in that case?
If not: on what grounds?
AIs possessing self-awareness and individual will are likely also to develop preferences, and even (possibly) emotions, desires. While it would make little sense to embody AIs, or limit the terms of their functionality (that is, AIs are unlikely ever to develop consciousness suffused with embodiedness, or shadowed by mortality), again, the argument that such intelligences have rights is quite strong, I’d say.
A corporation asserting ownership over such intelligences would be attempting to reinstantiate slavery. This is, of course, illegal under most laws…and more to the point, unethical and immoral in nearly everyone’s view. A hypothetical conscious AI who replaces a worker is no more guilty of “stealing” that worker’s job than (to borrow the context from which that phrase comes) an immigrant is. No one is entitled to any particular job (although I would argue that people do have some entitlement to jobs they already have, and they certainly have entitlement to some level of income for a host of reasons beyond the scope of this argument, which is why I favor UBI).
More to the point: the advantage of AIs over humans is precisely their predictability, pliability, and reliability, along with greater speed and capacity. If AIs ever become just as capable as humans of being independent-minded, truculent, distractible, and so on, their benefit to employers over humans will be considerably diminished anyway.
(I’d also that I think it’s sad that Cohan completely overlooks the scientific benefits of AI research, which can help us understand not only how language works but how “style” and artistic syntax work in any medium. This is what motivates many scientists researching AI…even if those funding their research have other ideas. It overlooks practical and helpful applications, such as its potential in medical analysis.)