Licensing destroys small startup businesses – deters investment

Before turning to the legal issues, it is important to appreciate a few salient facts about AI technology and its development.

First, the only practical way generative AI models can exist is if they can be trained on an almost unimaginably massive amount of content, much of which (because of the ease with which copyright protection can be obtained) will be subject to copyright. For example, large language models are trained on something approaching the entire corpus of the written word.

Second, AI models are not vast warehouses of copyrighted material, and any suggestion to this effect is a plain misunderstanding of the technology. As a result, when an AI model is trained on copyrighted works, the purpose is not to store any of the potentially copyrightable content (that is, the protectable expression) of any work on which it is trained. Rather, training algorithms are designed to use training data to extract facts and statistical patterns across a broad body of examples of content—i.e., information that is not copyrightable.

Third, as an empirical matter, the overwhelming majority of the time, the output of a generative AI service is not “substantially similar” in the copyright sense to any particular copyrighted work that was used to train the model. Even researchers employing sophisticated attacks on AI models have shown extremely small rates of memorization. 17

Fourth, over the last decade or more, there has been an enormous amount of investment—billions and billions of dollars—in the development of AI technologies, premised on an understanding that, under current copyright law, any copying necessary to extract statistical facts is permitted. A change in this regime will significantly disrupt settled expectations in this area. Those expectations have been a critical factor in the enormous investment of private capital into U.S.-based AI companies which, in turn, has made the U.S. a global leader in AI. Undermining those expectations will jeopardize future investment, along with U.S. economic competitiveness and national security. Turning to the legal questions raised by the Office’s NOI, a16z believes that generative AI model training is a productive, non-exploitive use of training material. That type of use does not exploit any protectable expression in any given work, and so it does not implicate any of the legitimate rightsholder interests that copyright law seeks to protect. It is for that reason that model training falls squarely under the fair use doctrine, the purpose of which is to “avoid rigid application of the copyright statute when . . . it would stifle the very creativity which that law is designed to foster.” Stewart v. Abend, 495 U.S. 207, 236 (1990); see also H.R. Rep. 94–1475 at 65–66 (1976) (fair use should be “adapt[ed]” to account for “rapid technological change”). That conclusion follows a long line of established precedent. Where copies of copyrighted works are created for use in the development of a productive technology with non-infringing outputs, our copyright law has long endorsed and enabled those productive uses through the fair use doctrine. Without the safeguard of fair use, we could not have now-ubiquitous technologies like internet search engines, see Kelly v. Arriba Soft Corp., 336 F.3d 811, 818–22 (9th Cir. 2003), online book search tools, see Authors Guild v. Google, Inc. (Google Books), 804 F.3d 202, 217–18 (2d Cir. 2015), and video game emulators, see Sony Computer Entm’t, Inc. v. Connectix Corp., 203 F.3d 596, 603–08 (9th Cir. 2000). Each of these technologies involves the wholesale copying of one or many copyrighted works. The reason they do not infringe copyright is that this copying is in service of a non-exploitive purpose: to extract information from the works and put that information to use, thereby “expand[ing] [the works’] utility.” Google Books, 804 F.3d at 217–18.

For the very same reason, the use of copyrighted works en masse to train an AI model—by allowing it to isolate statistical patterns and non-expressive information from those works—does not infringe copyright either. Imposing infringement liability for the use of copyrighted works in AI model training, notwithstanding the case law that clearly demonstrates why such uses are fair, would be extremely misguided. Among other things, it would upset at least a decade’s worth of investment-backed expectations that were premised on the current understanding of the scope of copyright protection in this country. The United States is currently at the vanguard of the AI industry as a direct result of these expectations and investments. But AI is not just being developed here in the United States; for example, it is also being developed in China, which views AI not as a tool for the betterment of humanity, but as a weapon for greater authoritarian control and influence. There is a very real risk that the overzealous enforcement of copyright when it comes to AI training—or the ad hoc limitation of the fair use doctrine that properly protects AI training—could cost the United States the battle for global AI dominance.

The bottom line is this: imposing the cost of actual or potential copyright liability on the creators of AI models will either kill or significantly hamper their development. And, significantly, treating AI model training as an infringement of copyright would inure to the benefit of the largest tech companies—those with the deepest pockets and the greatest incentive to keep AI models closed off to competition. A multi-billion-dollar company might be able to afford to license copyrighted training data, but smaller, more agile startups will be shut out of the development race entirely. The result will be far less competition, far less innovation, and very likely the loss of the United States’ position as the leader in global AI development.