Category: Artificial Intelligence (AI)

The Campaign to Defend Generative AI

generative ai

I have not written steadily about AI and copyright because, frankly, it’s exhausting. Not quite as exhausting as watching the state of the Republic overall, but almost as relentlessly incoherent and repetitive. For instance, Winston Cho for the Hollywood Reporter describes a PR and lobbying campaign by the tech coalition Chamber of Progress to defend the importance of generative AI (GAI). The article quotes founder and CEO Adam Kovacevich thus:  “Gen AI is a net plus for creativity overall. It’s expanding access to creative tools for more and more people and bypassing a lot of the traditional gatekeepers.”

That GAI may yield some beneficial tools for creators is plausible, but the whole “access” and “gatekeepers” rhetoric is a misguided anachronism from a group calling itself the Chamber of Progress. Perhaps “Confederacy of Tech Overlords” was too on the nose, but the generalized argument that GAI represents a “democratic” shift away from gatekeepers, stands on the rubble of experiments that have already failed. I doubt there is a professional creator left who hasn’t figured out that Big Tech’s promise to liberate them from traditional gatekeepers is like a human trafficker promising his next victim a job in a foreign country. Whatever was imperfect about the old models, the new models are more exploitative and hazardous for the average creator.

More precisely, while the alleged “liberation” from older distribution channels might have seemed attractive, GAI is about production, and I am confused as to who the “gatekeepers” would be on the production side of the equation. To the extent, say, Midjourney might enable me to illustrate or paint without any drafting or painting skills, the “gatekeeper” is who exactly? Nature failing to gift me with those skills? Or if we think big, and I can make a whole motion picture without ever turning on a camera, I still fail to see who the “gatekeeper” is in the overreaching promise from the tech industry.

Despite how cutting-edge and “essential” GAI is supposed to be, Big Tech has nothing fresh to say in its advocacy. The theme of “democratization” is the same weather-beaten argument they’ve been flogging for years, one that has proven disastrous for information and the state of real democracy—and which GAI can only make worse. Nevertheless, the Chamber of Progress campaign, as reported by Cho, seeks to promote a sweeping policy that AI developers should be broadly shielded from liability, including copyright infringement claims.

The question of copyright infringement for ingesting works for machine learning (ML) is currently at the heart of several lawsuits. I’ve lost track of them all, but arguably the most solid claim to date is New York Times v. OpenAI et al. because the evidence of copying (i.e., that what went into the model came out of the model) is so compelling. On the other hand, it is worth watching those cases where “reproduction” is less evident and, therefore, where the question may be more thoroughly addressed as to whether ML is a purpose that favors fair use of protected works.

As we have seen in defense of social platforms, Big Tech will spray the blogosphere with the term “fair use,” and copyright antagonists (mainly in academia) will echo the broad claim that of course ML is fair use. Notwithstanding the bugaboo that the fair use doctrine rejects the notion of a general exemption, I would argue that the case law points the other way, including the Supreme Court decision in Andy Warhol Foundation v. Lynn Goldsmith. To the limited extent that opinion addresses the ML question at all, its reigning in of the “transformativeness” test is more likely to disfavor the AI developers. Big Tech’s claim is that GAI is broadly “transformative” as a technological accomplishment, but Warhol and other decisions reject such a sweeping interpretation of that aspect of fair use factor one.

Further, as argued in this post, I remain unconvinced that GAI necessarily advances the purpose of copyright to promote new authorship as a matter of doctrine. For instance, if a given work created by GAI cannot be protected by copyright, then the material is, by definition, not a work of “authorship.” As such, this purpose should doom a fair use defense, in my view. Regardless, Big Tech will not be satisfied with the outcomes of any lawsuits, even if the developers win some. What they want is blanket immunity for infringement liability and an affirmation that GAI is truly as important as they say it is. That’s why this paragraph in the Hollywood Reporter story caught my attention:

In comments to the Copyright Office, which has been exploring questions surrounding the intersection of intellectual property and AI, Chamber of Progress argued that Section 230 – Big Tech’s favorite legal shield – should be expanded to immunize AI companies from some infringement claims.

Why highlight that? Because the absence of legal foundation is telling. Not only does Title 47 Section 230 have nothing to do with copyright infringement, but both that law and its copyright cousin, Title 17 Section 512, address the subject of users uploading material to platforms. Neither law says anything about scraping the web to feed material into an AI model for the purpose of ML. Nevertheless, it is clear from reading the actual comments by Chamber of Progress to the Copyright Office that Big Tech recommends policymakers take lessons from both statutes to carve out new liability shields to support the advancement of AI.

Despite the fact that neither §512 nor §230 has proven effective in limiting copyright infringement or dangerously harmful material online, the Chamber of Progress comments reprise Big Tech’s unfounded talking points regarding both statutes. Written by counsel Jess Miers, the comments repeat the false allegation that §512 fosters rampant, erroneous takedowns and also argues that because of §230, “most UGC services go to great lengths to proactively clean-up awful content and provide a safe and trustworthy environment for their users.” Not only will my friends and colleagues fighting Image-Based Sexual Abuse, online hate, and scams be very surprised to learn that, but so will Congress.

One of the scant points of agreement on Capitol Hill these days is that lawmakers have grown weary of liability shields for Big Tech, which has done a poor job of mitigating the worst harms facilitated by their platforms. Section 230 is so ripe for amendment that I’m surprised the Chamber of Progress invoked it, let alone in comments to the Copyright Office which only deals with, y’know, copyright law. More broadly, though, when GAI implies myriad harms beyond copyright infringement, the last thing Congress should do is grant Big Tech more latitude to do whatever it wants in the name of “progress.”  We tried that approach. It sucks.

Decoder podcast: AI could go extinct because fair use is whimsical

AI extinction

It was hard not to dismiss the headline posted by The Verge:  How AI copyright lawsuits could make the whole industry go extinct. The article summarizes a new Decoder podcast hosted by Nilay Patel, joined by Sarah Jeong to discuss the copyright lawsuits filed against generative AI developers. Most of the program is devoted to a discussion of fair use, which is reasonable because that’s likely how these cases will be decided. It’s clear that Patel and Jeong view copyright as a barrier to technological innovation, but when people trained in the law misrepresent the law as purely whimsical, it is counterproductive to the conversation.

I could critique nearly every segment in the podcast, but as that would be both long and tedious, I selected a few highlights for this post. Setting the more-hip-than-helpful tone of the program, Patel (who went to law school) describes fair use as a “vibes based” doctrine. Jeong (also law school) echoes the sentiment when she says that litigation against generative AI has “Napster vibes to it,” teeing up her thesis statement: “When Napster happened to the law, companies went bust; entire industries went bust; copyright changed forever in a way that was not great; it was an extinction level event; and AI has a similar thing going on there.” Here, Patel summarizes that Napster went to the Supreme Court—it did not—and that the Court “made some changes to copyright law.” Seriously? “Made some changes” is not how people with legal training talk about court rulings, even when they disagree with the outcome.

The next comment that caught my attention was Patel saying that “fair use is not deterministic” as a doctrine. He’s right, but in context, the listener will take him to mean that fair use is unpredictable to the point of capriciousness. Although a good attorney will demur to predict the outcome of any case, a thoughtful copyright expert is unlikely to agree that fair use findings are a “coin toss,” as Patel puts it. In fact, the choice of the word deterministic provokes the rebuttal that anticipating a fair use outcome is more accurately described as probabilistic, which is funny because that’s also how generative AI works.

If a defendant asks an attorney to handicap the likelihood of prevailing on fair use, the attorney’s response should be a reasonable prediction based on how closely the facts of the present case resemble fair use findings in the circuit of jurisdiction. Although Patel alludes to this analysis, he overlooks the fact that counsel could describe a probability outcome, which is precisely how a generative AI produces its outputs. If one prompts a visual AI to generate an image of a dolphin drinking a Slurpee, the output is the machine saying, “Based on the available data, this image is probably a dolphin drinking a Slurpee.” So, of all defendants, AI developers should grasp the nature of fair use case law.

Jeong echoes the idea that fair use considerations are erratic by alleging that the “Court changed copyright law after Napster,” referring to the Ninth Circuit’s 2001 finding that the P2P music filesharing platform was not shielded by fair use. Here, she argues that the Supreme Court’s fair use finding in the Sony “Betamax” case (1984) expressed a philosophical adaptation of copyright law to foster new technology but that this general view was reversed when the Ninth Circuit decided against Napster—and then when the Supreme Court ruled in 2005 that the filesharing platform Grokster could be liable for copyright infringement.

Although one cannot reasonably argue that ideology never skews the courts, Jeong elides the many factual and legal distinctions between the VCR and filesharing platforms and, by extension, the distinctions between those technologies and generative AI. Her declaration that “copyright was changed” after Napster and Grokster is unfounded, as the Court itself notes that Grokster was its second case considering contributory liability for copyright infringement—Sony being the first. Two cases, twenty-one years apart, addressing the same legal question presented by substantially different technologies is not a basis for claiming that the law was “changed forever” by the outcome in the latter case.

Holding the opinion that copyright stifles technological innovation does not excuse misrepresenting the courts as rolling dice to rule on fair use. For instance, in Grokster, the Court directly addresses the balance between copyright and technological innovation thus:

The more artistic protection is favored, the more technological innovation may be discouraged; the administration of copyright law is an exercise in managing the trade-off….The tension between the two values is the subject of this case, with its claim that digital distribution of copyrighted material threatens copyright holders as never before, because every copy is identical to the original, copying is easy, and many people (especially the young) use filesharing software to download copyrighted works.

Does that describe the technological function of the VCR? For those who’ve never used a VCR, the answer is No. The home video tape recorder, a relic of pre-internet life, functioned nothing like a filesharing platform, which facilitates mass copyright infringement on a global scale. Fair use is a fact-intensive inquiry, and “technology” is not a monolith. The leap from the VCR to generative AI is roughly the distance between the telegraph and the iPhone, and it is unhelpful, even irresponsible, to obscure so much factual detail behind a conversation about the courts’ alleged randomness on copyright and fair use.

Everything cited above was expressed in the first 5-6 minutes of the podcast. Tempted not to listen any further, I winced as both Patel and Jeong proceeded to make some astonishing remarks about the four-factor fair use test in regard to generative AI. Again, a couple of highlights stand out.

On factor two, nature of the work used, Patel says, “Factor two is whatever the judge thinks it is.” Then, a few seconds later, he says, “If the judge decides they don’t like the New York Times that day…” this will determine whether factor two tilts in the Times’s favor. NYT v. Open AI is before the Southern District of New York in the Second Circuit, which holds the largest trove of copyright case law of any circuit in the country—including several major fair use cases. If Patel or Jeong want to handicap the court’s findings based on that case law and then offer their own views of what they think is right, fine. But the implication that the court is just going to wing it is ridiculous.

Jeong does not push back on Patel’s coin-toss implication but says the “dial is in the middle” on factor two, which she reasonably (if not very clearly) argues because the Times contains both protectable expression and unprotectable factual material. But then comes the biggest spit-take in the program, when Jeong predicts that factor four, potential market harm to the work used, weighs against the AI developers because of the Supreme Court decision in Warhol. She states, “We have not seen that heavy an emphasis on factor four before.” Notwithstanding the fact that prior to the Campbell decision (1994), many experts would say that factor four was the most determinative factor in fair use jurisprudence, Warhol was unequivocally NOT a factor four case. As the opinion states:

In this Court, the sole question presented is whether the first fair use factor, “the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes,” §107(1), weighs in favor of AWF’s recent commercial licensing to Condé Nast. [emphasis added]

From her comments about Warhol, Jeong confuses the question of “substitutional purpose” with the question of “market harm substitution,” which are weighed under factors one and four respectively. It is true that where the court finds substitutional purpose, market harm substitution is more likely to be found, but as the opinion explains the distinction in one footnote in Warhol:  While the first factor considers whether and to what extent an original work and secondary use have substitutable purposes, the fourth factor focuses on actual or potential market substitution. They are two separate, albeit interdependent, questions. I do not know whence Jeong gets the idea that Warhol was a factor four case, let alone an unprecedented outcome in its emphasis of that factor.

Fair Use is a Fact-Intensive Inquiry

Generalizations like those articulated in the Decoder podcast sidestep the relevant facts about a given technology, what it does in context to legal questions, and why the technology may or may not be socially valuable. “Artificial intelligence” encompasses a wide range of development, some of which is promising, some of which is questionable, and all of which has been identified as potentially dangerous without proper oversight. As for generative AI in the creative industries, if Jeong is right that the copyright lawsuits pose an existential threat to those companies, so what? It is not clear that the world needs machines to make images of dolphins drinking Slurpees.

As discussed in this post, AI developers may have taken a gambler’s approach to fair use, and if their business plan included liability at the scale of mass copyright infringement, that’s a risk they chose to take. If any of those companies fail because of that liability, it will not be the result of whimsically applied or tech-hostile copyright law, or indeed the fault of the creators whose rights are infringed in the process of machine learning. Moreover, it is certainly not incumbent upon creators to abdicate their rights and get out of the way because “innovation” is happening. Fair use considerations in generative AI lawsuits may result in some novel opinions, but if influencers like Patel and Jeong are going to misstate case law and describe the courts as casinos, then one must wonder why they mention their legal credentials in the first place. After all, anyone can flip a coin.

Stop Democratizing Everything!

democratizing

On March 17, Rolling Stone published an article featuring a song called “Soul of the Machine.” Sounding like blues of the early 20th century, the “voice” sings the lyric, “I’m just a soul trapped inside this circuitry.” Naturally, the whole work—music, lyrics, guitar playing, and singing—was produced by artificial intelligence. As writer Brian Hiatt describes, a simple prompt, “solo acoustic Mississippi Delta blues about a sad AI” produced the song after a fifteen-second collaboration—music and performance by Suno with lyrics by ChatGPT. Yes, it’s a “Holy shit” result with a million implications, but it was this paragraph about Suno’s co-founder that inspired today’s response:

Suno appears to be cracking the code to AI music, and its founders’ ambitions are nearly limitless — they imagine a world of wildly democratized music making. The most vocal of the co-founders, Mikey Shulman, a boyishly charming, backpack-toting 37-year-old with a Harvard Ph.D. in physics, envisions a billion people worldwide paying 10 bucks a month to create songs with Suno. The fact that music listeners so vastly outnumber music-makers at the moment is “so lopsided,” he argues, seeing Suno as poised to fix that perceived imbalance.

At some point—and I think it’s the point on top of most technologists’ heads—the word democratization became a handy euphemism for destruction. Social platforms “democratized information,” and we’re drowning in disinformation. Streaming platforms “democratized distribution” for creators and decimated royalties. And now, generative AI developers want to “democratize creative production” with the snake-oil pitch that everyone can be a painter, musician, filmmaker, poet, etc., as if art is something to heat up in the microwave like a quick (if not good) meal.

The first rule of economics is that abundance lowers value, and this does not only apply to price but also to those esoteric values we ascribe to the artistic works that attain meaning for us. In Shulman’s view, Bob the electrician would “make” his own big band music while Sally the paralegal would “make” her own Reggae, and if we multiply that to the scale Shulman projects above, then a billion people can “make” music about which a billion people do not give a damn. Consequently, as argued in this post in January 2023, the inevitable outcome of this entire enterprise is widespread boredom.

It is not possible to “democratize” the production of art in the way Shulman envisions because the individual who types a few words into an AI to produce a “new” song will never experience anything close to the process of making music. As described by Hiatt, the “production” of “Soul of the Machine” is the equivalent of saying, “I’m in the mood to listen to Mississippi Delta blues,” which describes how most of us decide what to play at a given moment. But that’s not making music, it will never feel like making music, and few people will ever feel otherwise.

I can’t play guitar for shit, but because I am a human being composed of human parts, I sense the extraordinary degrees of difference between listening to Mark Knopfler and trying to force my lame-ass fingers to make those sounds. As such, it would take a traumatic brain injury for me to be deluded enough to feel like typing a prompt to direct a machine to play a Knopfler-like solo was somehow an accomplishment in this regard. Artistic works need to be special, and whatever makes them special also needs to be a shared human experience for the work to matter. Lacking these ingredients, “art” produced by a machine is just a Hot Pocket in the microwave.

When I first jumped into this fray, EVERYBODY on the anti-copyright side was preaching to creators that they need to forget about “old models” built on sales and royalties and instead embrace online platforms to “connect to their fans.” Follow this new doctrine, they insisted, and fans will reward them as a courtesy rather than be forced to pay “rent” by a government-imposed monopoly called copyright. Yes, it was multi-dimensional bullshit ten years ago, other than the fact that certain creators could, and can, connect with fans in novel ways. But now, the same class of tech-bros, heavily invested in generative AI, propose to wipe out that connection with the new promise that today’s fans are tomorrow’s artists.

I get how Suno makes a good pitch. An addressable market of a billion people paying ten bucks a month is going to get VC attention. But like all utopian “visionaries,” generative AI developers’ dreams of “democratizing” creative production forget to consider human nature, without which art is meaningless. After the initial gee-whiz factor wears off, the music or writing or painting itself all amounts to a big Who cares? “Soul of the Machine” is an impressive, eerie accomplishment in computer science—one that will doubtless have applications—but if we proposed to send a new Voyager mission beyond the solar system with a new gold disk telling a human story, Blind Willie Johnson would still belong, and not some probability outcome produced by a generative AI.

Meanwhile, I still wonder whether the model itself might crash as its own self-training approaches a state akin to consciousness. The lyric about being trapped inside the circuitry is satire for humans that reprises a question I’ve asked before—namely whether an AI might attain semi-consciousness and begin to produce what it perceives as “art.” Specifically, the question is whether the AI might ever “understand” its nature and then make expressions about the “machine condition” rather than randomly produce ersatz expressions about the “human condition.”

While I am told by some technologists that this idea of near consciousness remains in the realm of science fiction, my own bias still predicts that if the AI could ever ask itself why it should produce art, it probably won’t. Or if it does, it will be in the form of expressions that we would not understand—or perhaps even know exist. So, even if Shulman’s “boyishly charming” vision were achieved at some scale, I predict it will start to suck, and suck fast. Then, like a reverse Fahrenheit 451, as the over-abundance of bespoke music threatens to burn the old catalogs out of living memory, people will “rediscover” the real thing, and the proverbial children in the woods will know the difference.

Photo by: Talulla