The New “Normal”

Paul Ohm

For those splashing around in the shallow water of AI Law and Policy, it’s a heady time to be mapping old laws onto transformative AI-based technologies, such as chatbots, image generators, and AI agents. But off toward the horizon in the deeper water are worrisome shadows, forecasts for the arrival of Artificial General Intelligence (AGI), and confident but contradictory predictions about how AGI will lead to either tech Nirvana or human extinction (and sometimes, confusingly, to both). Bobbing around out there are adversarial flotillas firing potshots at one another, flying obscure banners with confusing tribal names: e/acc, long-termism, TESCREAL, to name only three. The rational thing to do has been to cover one’s eyes and ears, trying to block it all out.

There is now a third and better option to either avoiding or diving deeply into the AGI waters: read the recent works by two Princeton computer scientists, Arvind Narayanan and Sayash Kapoor, who have started to engage with and rebut the most radical calls to action of the AGI-set. To get the full extent of their rich argument, you should read not only their recent book, AI Snake Oil, but also their subsequent series of online essays, some as long as traditional law review publications. Only by considering four different works together (and one imagines there will be more to read from the pair soon), does the full argument take shape. The good news is that their argument—if it is correct—dispels worst-case predictions about both the speed with which the problems of AGI will emerge and the radical nature of what we will need to do to respond to those problems.

AI Snake Oil is two books in one, and the title describes only one of them. When it comes to pre-generative, non-foundation-model AI—the models hawked to criminal justice systems to predict whom to incarcerate, employers to decide whom to hire or fire, and universities to decide whom to admit—Narayanan and Kapoor document a world awash in snake oil. Opportunistic tech hucksters sell overhyped services to nontechnical procurement experts at underresourced and vulnerable (they call them “broken”) institutions, contributing to inequity, unfairness, and human misery. For those who are coming to AI Law and Policy anew (including law students), this half of the book will provide an invaluable, crystal clear introduction.

Interwoven with these stories of snake oil, however, is a book that is surprisingly bullish about the transformative power of another kind of AI: the powerful new generative and foundation models, such as large language models, powering AI chatbots and agents. Although Narayanan and Kapoor are very worried about some harms that these models are causing, they are at the same time quite optimistic about the potential these models have for transformative good. They seem not to consider the companies creating and fine-tuning these models to be snake oil sellers. It’s confusing that the book is saddled with a title that fails to describe half of it; one sees the uncompromising hand of the publisher’s marketing department.

The most important move of the book—as extended in the three posts—is a detailed argument against dire predictions offered by many that the rise of so-called Artificial General Intelligence (AGI) risks catastrophic harms up to and including extinction of the human race. Narayanan and Kapoor make two very important moves in response. Both of these moves appear in inchoate form in Chapter 5 of the book, but are given richer shape in the subsequent essays.

Move one is about the timeline for AGI. Narayanan and Kapoor divide the rise and spread of powerful technologies into three distinct acts: invention (the basic scientific or engineering act of creation), innovation (the adoption of an invention into a useful tool), and diffusion (the “broader social processes” that lead to the widespread adoption of the tool). For example, creating the next version of a large language model is an invention; using this LLM to power a new and improved legal research tool is innovation; and convincing law firms to adopt the tool en masse is diffusion. They argue that most of the generative AI advances we have seen to date count as invention, with a smattering of innovation; diffusion is as yet a distant goal.

They argue, in AI as Normal Technology, that before we can ever get to AGI or the concomitant risks of catastrophe, much more innovation and diffusion will need to happen. This won’t happen overnight, because innovation and diffusion require slow and messy society-wide, human-bound, friction-full advances (which they describe through the helpful metaphor of climbing a “ladder of generality”), the kind you can’t just speed along by purchasing more chips from Nvidia.

Move two is more important for legal scholars, although it gets a bit underplayed in these texts: trying to prevent the catastrophic harms of AGI at this early stage, such as through new AI safety laws, would be premature. Anything we would do now would aim for the wrong targets and solve the wrong problems. They offer this analogy in the original AI as Normal Technology essay:

At the dawn of the first Industrial Revolution, it would have been useful to try to think about what an industrial world would look like and how to prepare for it, but it would have been futile to try to predict electricity or computers. Our exercise here is similar. Since we reject “fast takeoff” scenarios, we do not see it as necessary or useful to envision a world further ahead than we have attempted to. If and when [AI models achieve “superintelligence”], we will be able to better anticipate and prepare for whatever comes next.

In other words, even if the existential risks of superintelligence will be upon us at some point in the future, anything we do today to prepare for them will turn out to have been mistargeted when the time comes. To be clear, the authors are arguing against only what they see as overly aggressive forms of AI safety laws, citing the 2024 Frontier AI safety law that Governor Gavin Newsom vetoed as an example. Their argument does not extend to laws that address the here-and-now harms of the generative AI we have today and the newer, more powerful versions we will have soon, such as misinformation, deepfakes, labor displacement, or environmental impacts.

Narayanan and Kapoor make a particularly important contribution in the clever label they have coined as shorthand for their analysis: AI as Normal Technology. This has the potential to be the next “Code is Law” or “Surveillance Capitalism,” the bumper-sticker-worthy catchphrase for disaffected legal scholars and policy wonks everywhere. Consider a few defining features of this phrase.

First, AI as Normal Technology will lead to constructive misinterpretation. Where do Narayanan and Kapoor get off calling generative AI “normal?” The authors themselves have explained that AI’s normalcy is in relation to the extraordinary claims being made by the most science-futuristic of AGI’s believers. AI is normal alongside other “normal” technological revolutions, such as electricity and the internet, two significant and transformative markers in the history of technological development.

Second, the phrase can contain multitudes. AI is significant and disruptive and perhaps transformative, but it is Normal. We can control it through normal order, not by acting like nothing like this has ever happened before. From race scholars inveighing against AI bias, to the new “librarians of Babel” engaging in acts of conscientious objection on college campuses, to state legislators advocating against federal preemption, the idea that AI is “Normal” Technology is a rhetorical shrink ray. It brings grandiose and overwrought claims back down to their appropriate size. All of the AGI doomers are just engaged in ordinary acts of persuasion and politics and lawmaking and advocacy. The Normalcy of “AI as Normal Technology” erects a very big tent, giving those sheltered by it a strength in numbers.

In fact, this tent is so big that it covers many who might disagree with important parts of what Narayanan and Kapoor are arguing. To my mind, for example, they focus too much on resiliency as a catch-all strategy for dealing with AI safety harms. I would have preferred that they focus more on the way AI will reinforce the power of the already too-powerful, in other words, on the political economy that will make it difficult to prevent AI from being used to do anything that does not consolidate market share, control, and revenue in a few tech oligarchs and their political puppets.

Deep down, I think Narayanan and Kapoor agree with me on some of this—the main takeaway of the entire book is, “we are not okay with leaving the future of AI up to the people currently in charge”—but this storm-the-barricades fervor seems to recede in their later essays, which center on milder interventions such as increasing funding for research on AI risks, developing early warning systems, and promoting competition. The obvious reason AGI might lead to runaway harm is that a small number of investors will be profiting from it; we can and should address this root cause today.

The third, and perhaps most important, work done by the AI as Normal Technology framing is that it places the burden of proof where it belongs: on those arguing that the next few years and decades of AI will be unlike any other technological development in the history of humanity. Many serious and intelligent people have embraced these arguments. The “Normal technology” frame serves to remind them—and the decisionmakers trying to make sense of their arguments—that they are the ones advancing the exceptional, the unprecedented, indeed, the abnormality of our changing world. They might in the end turn out to be right, but I am not betting on it. We have seen a lot and survived a lot, and viewing where we are as normal order seems the more reasonable bet.