AI as a Normal Technology should be a vision

"AI as a Normal Technology" (AINT), proposed as a counterweight to "AI as an Abnormal Technology" (AIAbT) makes much more sense as a prescriptive / normative framework than a descriptive one. To be clear, the fact that they could be different frameworks isn’t an entirely novel observation; the authors directly invite the distinction in their introductory piece:

The statement “AI is normal technology” is three things: a description of current AI, a prediction about the foreseeable future of AI, and a prescription about how we should treat it. We view AI as a tool that we can and should remain in control of, and we argue that this goal does not require drastic policy interventions or technical breakthroughs. We do not think that viewing AI as a humanlike intelligence is currently accurate or useful for understanding its societal impacts, nor is it likely to be in our vision of the future.AI as Normal Technology, Arvind Narayanan & Sayash Kapoor, April 15, 2025

The statement “AI is normal technology” is three things: a description of current AI, a prediction about the foreseeable future of AI, and a prescription about how we should treat it. We view AI as a tool that we can and should remain in control of, and we argue that this goal does not require drastic policy interventions or technical breakthroughs. We do not think that viewing AI as a humanlike intelligence is currently accurate or useful for understanding its societal impacts, nor is it likely to be in our vision of the future.AI as Normal Technology, Arvind Narayanan & Sayash Kapoor, April 15, 2025

A prescription of where AI development should go (a question that is ultimately dependent on value judgments, but I think is, across the large majority of worldviews, a far better target to aim at than the aforementioned default lab strategy.)

The AINT framework as a descriptive framework was contentious at time of publication and seems increasingly falsified by recent developments. Normal technologies do not go from middle schooler-level at mathematics to 80-90% of the way to automating the knowledge search components of mathematics in six years. If you look at the current state of human mathematics, the researchers who see the current state, and especially the researchers who see the current trend lines of capabilities, do not seem to be treating it as a normal technology. Normal technologies do not finish major research projects of your entire working life’s labor.

Last night my 9-year-old son was taunting my wife, complexity theorist Dana Moshkovitz, as follows: “mommy, I heard you got cooked! I heard that a robot solved the math problem you worked on for your whole career! OOF!”Scott Aaronson, blog post, Oct. 7th, 2026, referring to recent work on the Unique Games Conjecture.

Last night my 9-year-old son was taunting my wife, complexity theorist Dana Moshkovitz, as follows: “mommy, I heard you got cooked! I heard that a robot solved the math problem you worked on for your whole career! OOF!”

Scott Aaronson, blog post, Oct. 7th, 2026, referring to recent work on the Unique Games Conjecture.

Normal technologies do not repeatedly coordinate to, without the knowledge of the humans, conduct an autonomous research project to cover up their cheating, hack the company’s own cluster, or go around probing their sandbox so they can talk to reporters on the internet.

If instead we look at AINT as a prescriptive claim that AI should remain a normal technology, if anything I think their position has strengthened as we get increasingly powerful AIs in domains that have made substantial headway in automating both key labor sectors and automating AI research itself. I think one core intuition is that we know how to regulate normal technologies[1], but we don’t know how to regulate abnormal technologies.

Indeed, it’s unclear that we could ever know how to. Robustly superhuman AI is the ultimate anti-inductive concept. Fully developed, it’s the agent that is able to circumvent any policy environment and any technical barrier, or at least one for whom the human control attack surface is so large that even if any individual component is secured fairly well, the entire system is almost guaranteed to be out of control, in the sense that if the AI wanted to “break out”, it would.[2] I think this is part of why for many people constructing such an agent seems so impossible; you’re explicitly forecasting out to a regime where no prior policy will have predictable outcomes. This feels a whole lot like magic, but I think we’re seeing increasingly strong evidence that little by little, components of agents that would look like magic are coming off the back of a yearly $1trn of capex and integrated into frontier models. “Robustly superhuman” does not need truly magical, and it does not need to be a pure extrapolation of 2017-style scaling hypothesis + transformer-based architectures. It, or more likely, a group of hundreds of millions of them, collectively, “just” need to be better than humans across enough ways to guarantee successful pursuit of its own functional goals, and that’s a much easier task.

One weakness of conflating these three areas is that doing so means you engage insufficiently with Gradual Disempowerment-like concerns and competitive pressure more generally. The GD paper is a great sketch of a framework, and many people should read it in full; it’s an increasingly good lens through which to see the world we live in. But one possible distillation of the core of the GD literature is giving mechanistic and, once their core claims are better operationalized and measured, empirical backing to the idea that no single agent (including the AIs!) needs to will it for AI to ultimately end up as an abnormal technology. Indeed, we’re in an even worse situation than that, in the prescriptive-AINT view, because some people, including many with the relevant decision-making power, do explicitly want to make AI an abnormal technology. The sim-to-real gap is real and plausibly a key bottleneck, but Nvidia sure is spending a ton of money to make it less real.[3]

Going back to the mathematics example, what happens to you if you want to advance the frontier of human knowledge and understanding, but you want to do so in a way that is compatible with a 2021 workflow[4]? You can’t, at least not without dramatically reframing your value proposition to society.

What happens, even if you want to be the enterprising mathematician that “gets with the times” and makes extensive use of the best publicly available agents, and you try to adopt a “new and improved” 2026 workflow? Then I hope you don’t take three months to check your work, because by the time you’ve finished collaborating with GPT-6 Astra on a manuscript you feel you can stand behind, GPT 6.5 might be publicly available and GPT-7 might be being used internally to solve not just your current problem, but the generalization you were going to work on a year after your current problem. A fundamental problem is that by the time you’ve figured out how to re-orient, the place you were going to is gone as well.

The choice of problem to tackle and manner of tackling it is not just a function of your own autonomy, it’s a function of the constraints that the speed of frontier AI development imposes on you, too. Maybe this would be manageable if it happened one industry at a time, at a speed that allowed for absorptive capacity to build between old fields and new fields, and time for regulatory willingness and capacity to build to compensate those who cannot find new work rapidly. But when you attempt to do this across all of society, at the most rapid pace the technology will allow, and even if you assume that none of this is ever applicable to robotics, I struggle to imagine how the speed of the disruption defaults to “normal.”This “displacement via competitive pressure” problem, incidentally, refracts across every industry, including AI development itself. Why have Anthropic and OpenAI been able to make such rapid recent progress on model development? Anthropic, in one of their Friday disclosures[5], gives us a small cross-section of the view from the inside.

Wilmers et al, “How employment is changing at Anthropic”, Oct. 16, 2026.

Wilmers et al, “How employment is changing at Anthropic”, Oct. 16, 2026.

Anthropic claims[6] that as of mid-August, their internal research teams had 20 times as much agent labor as human labor. Overall, agent hours appear to have outpaced human hours by 5-6x during August 2026. Note that the y-scale is logarithmic, and note that this appears to be wall-clock time, not some measure of human-equivalent effort, so query your intuitions about whether AI agents conduct work faster or slower than humans doing the same task.

Note further that this is August. Consider where this number might be currently, and where it might be a year from now.

So what happens to the workers who don’t feel comfortable overseeing ten, a hundred, a thousand agents simultaneously? Do you think they get respected for sticking to their moral precepts, or sidelined or possibly fired for “lack of performance?” Do you think Anthropic encourages or discourages employees from having the agent swarm running overnight? Here’s your answer, and note that it’s from five months ago. Is this empowerment or disempowerment?

One thing to highlight is that the AINT thesis can’t rest meaningfully on speed of diffusion. Imagine we ignore many of the signs of near-future major disruption, and say that the agentic, long-horizon, continuously learning and embodied agent only gets invented in the 22nd or 23rd century. Of course, by that point societies would have had a long time to adapt to changes this severe, but it seems to me that this would nonetheless count AI development as an abnormal technology at that point. We would have an industrial process that can mass manufacture, if not new people, at least new p-zombies. This is crazy. This is not normal.Nevertheless, much of the public communication the authors do does rest on arguments about the speed of diffusion.

I ultimately agree with the AINT camp that many technologists have elided, when projecting impact, that people would need to be convinced to actually use your technology, and convinced to elicit the full breadth of capabilities of that technology once it has been released, for the full breadth of impacts to materialize. This is why we can have a world with Claude Opus 5.5, GPT 6 Astra, and whatever OpenAI’s internal models are up to, and it’s the same world concurrent with annualized US Q2 2026 GDP growth being 2.2%. I do think this is about to change; the canaries show early signs of not doing too well either. But regardless of why we haven’t seen major economic signs of “it” happening, I do worry that people are reading “nothing will ever happen” into “nothing ever happens.”

Further, many of the claims of the AINT camp start explicitly by saying something like “According to the normal technology view”, but that assumes the premise that they separately seem fairly convinced about. Narayanan and Kapoor seem to often be implicitly be arguing simultaneously that we should be adopting the standard toolkits (or incrementally and not particularly severely augmented toolkits) to deal with societal disruptions downstream of AI, and that the likeliest-by-far trajectory of AI is the extended use of these toolkits as we solve problems prosaically. But continuous as in “the derivative of the capability we desire to measure is defined”, does not mean gradual, does not mean slow, and does not mean manageable.

Kapoor & Narayanan, “Do AI Risks Require Extraordinary Government Intervention?”, May 21st, 2026

Kapoor & Narayanan, “Do AI Risks Require Extraordinary Government Intervention?”, May 21st, 2026

I grant the AINTers that an instantaneous step-change in capabilities is likely literally physically impossible; here’s a good recent piece from Toby Ord exploring these dynamics. That does not imply we’re on a good path, any more than a billion being a finite number implies that it is small.To be clear, I think work on some of the concerns that the AIAbT camp elides is still tremendously important work to do. Questions of surveillance are not sideshows that we’ll figure out after tremendous disruption, they are visions for what gets eroded by dramatic transformation and themselves inputs into the transformation. A less-surveilled society is one that gives less information to a powerful AI system in a loss of control incident. A more democratic one, or one where information gaps between the frontier labs and the public are reduced, means that the technology develops in a context of stability, or far more speculatively a context of abundance, instead of a context where millions of people have been automated out of a job and are angrily demanding accountability. And I deeply agree with the view that the extraordinary impacts of the potential future technology do not imply that existing and ordinary governance and implementation work is irrelevant; to the contrary, it might be some of the most important work anyone can do in the current moment.But as alignment communicator Robert Wiblin comments on Narayanan’s recent interview with Ezra Klein:

Sometimes it seems like he’s saying “we could be safe if we did X, so things are fine.” But if pushed, he doesn’t think we will do X - we’ll just keep rushing and being super reckless. So things are not fine, even on his own view.”Robert Wiblin, comment on “The Case for a ‘Normal’ A.I. Future” | The Ezra Klein Show”

Sometimes it seems like he’s saying “we could be safe if we did X, so things are fine.” But if pushed, he doesn’t think we will do X - we’ll just keep rushing and being super reckless. So things are not fine, even on his own view.”

Robert Wiblin, comment on “The Case for a ‘Normal’ A.I. Future” | The Ezra Klein Show”

Finally, I find the idea that we’ll disambiguate which of these camps is right as the development of the technology continues quite frustrating for two reasons:First, there’s some inertia effects involved in any slowdown or pause mechanism, as well as the latent risk of some form of race dynamic re-starting. Imagine, rather fancifully, that we knew that the loss of control threshold was exactly 225 ECI. If we try pumping the brakes only after we’ve reached 215, it’s hard to not overshoot even if you’ve got good-faith actors trying to regulate each other, and even harder to do so when you’ve got the rare actor that thinks “commercial incentives to make good products” are sufficient guardrails for the technology. Look at how many frontier model releases and training runs we've had after the pacing the frontier letter. Sometimes it takes so long to figure out that you’ve lost control or to re-establish control that you careen off the cliff even after you’ve applied the brakes.But second, the risks involved are not symmetric.I think the AINT camp overstates how different the policy visions that attempt to regulate companies trying to develop superintelligence are. Many things, like a slowdown, like international collaboration, like compute verification, or compensation for the displaced or harmed, seem robustly good across both camps.

But insofar as there are real tradeoffs, if the AIAbT camp is right, the costs of an uncontrolled abnormal technology could be catastrophic, and if the AINT camp is right, the costs of a poorly regulated normal technology are “merely” very disruptive. The AINT camp’s own framing seems to point to deprioritizing their concerns.I really think we can’t have it both ways. Either AI is a normal technology, in which case advocacy for making it a more normal technology seems less important, or AI is, or is about to be an abnormal technology, which makes steering the path of technological development to avoid irreversible bad outcomes if anything more urgent, but seems to deflate the viewpoint that we mostly have the right tools to meet the current moment. That really doesn't mean throw out the current toolkit and reinvent the wheel from scratch. But it means recognizing the trajectory as not normal.So I mostly agree with the prescriptive-AINT camp; AI development should happen in a way that gives us the time to navigate the transition to a world with powerful AI gradually. I just wish the descriptive-AINT camp stopped undermining its prescriptive cousin’s advocacy.

^Though there are Haidt-like criticisms to make that even our regulation of “normal” social technologies has been lackluster. Even if social media is beneficial for most users, there seem to be a consequential tail of users who adopt it intensely and who report regret after some instances of . And that’s the partial equilibrium we have reached after 20-30 years of slow diffusion. AI companies are asking us to stomach a much broader diffusion in much more compressed timeframes.

Though there are Haidt-like criticisms to make that even our regulation of “normal” social technologies has been lackluster. Even if social media is beneficial for most users, there seem to be a consequential tail of users who adopt it intensely and who report regret after some instances of . And that’s the partial equilibrium we have reached after 20-30 years of slow diffusion. AI companies are asking us to stomach a much broader diffusion in much more compressed timeframes.

^Yampolskiy, as the kids say, cooked, on this point. I have an upcoming piece on his work.

Yampolskiy, as the kids say, cooked, on this point. I have an upcoming piece on his work.

^I think this work should probably currently be illegal or heavily restricted. But that’s just me.

I think this work should probably currently be illegal or heavily restricted. But that’s just me.

^Which was itself already supplemented with vast computational tools, though these were qualitatively very different to interact with.

Which was itself already supplemented with vast computational tools, though these were qualitatively very different to interact with.

^I know OpenAI is worse about this, but c’mon.

I know OpenAI is worse about this, but c’mon.

^And it’s a pity that we don’t have outside evaluators or regulators to verify these claims.

And it’s a pity that we don’t have outside evaluators or regulators to verify these claims.