'AI as Normal Technology' and the case for boring

The claim

Arvind Narayanan and Sayash Kapoor published a long essay on April 15 at the Knight First Amendment Institute with a deliberately flat title, AI as Normal Technology. The word normal is doing three jobs. It is a description of AI today, a prediction that it will stay controllable and comparable to electricity or the internet, and a prescription that governments should treat it with the same sector-by-sector toolkit they use for everything else. The authors say up front that they have not tried to put probabilities on anything and that this is a first articulation rather than a full defence.

We read it because we spend most of our week with tools that feel anything but normal, and we wanted to see whether the argument survives contact with that experience. Most of it does. One piece of it, the pace of diffusion, is the part we keep turning over.

Three clocks running at different speeds

The essay's most useful move is separating invention, innovation and adoption and insisting they run on different clocks. Invention, meaning new methods, moves fast. Innovation, meaning products built on those methods, moves slower. Adoption, meaning organisations restructuring around the products, moves slowest of all. The example that stuck with us is from their own earlier work on predictive optimisation: high-stakes systems in criminal risk assessment and medicine still run on decades-old statistical models, because in safety-critical settings the cost of getting it wrong dominates the benefit of a newer method.

The historical anchor is electrification. Dynamos were, in the phrase they quote, everywhere but in the productivity statistics for close to forty years after Edison's first generating station, because factories had to be rebuilt around distributed motors before the gains appeared. Self-driving cars are the modern case. AlphaZero could iterate through self-play in hours. Cars needed more than two decades because each iteration had to be run on real roads under real safety constraints.

They pair this with a number we had not seen framed this way. By August 2024 around 40 percent of US adults had used generative AI, and the estimated productivity effect was 0.5 to 3.5 percent of work hours. Breadth of adoption and depth of adoption are different quantities, and benchmarks measure neither.

Where the argument is strongest

The section on misuse is the one we would hand to a policymaker. Their point is that model-level alignment is brittle for a structural reason. A model asked to write a persuasive email cannot tell whether it is being used for marketing or for phishing, because the information that distinguishes them lives outside the prompt. So defences have to sit downstream, in infrastructure, the way they do in security generally. They add that AI is also useful for defence, which means restricting development can hurt defenders more than attackers.

The systemic risk section is also better than the usual treatment. The Industrial Revolution comparison is not a comfort. They use it to point out that the disruption produced harsh working conditions, exploitation and inequality before institutions caught up, and they expect AI to concentrate power and erode trust in ways that need no superintelligence at all. That is a risk story you can act on with tools that exist.

Where our week disagrees with the essay

The diffusion argument assumes adoption is gated by organisational restructuring and by real-world feedback loops that are slow to run. Software is the domain where both gates are weakest. A coding agent's feedback loop is a test suite, which runs in seconds and costs nothing to fail. There is no regulator, no physical safety envelope, and the organisation being restructured is often one person and a terminal. The day after this essay went up, OpenAI released a terminal coding agent, two months after Anthropic released one, and both read, edit, test and commit. Uptake in the teams we know has been faster than any tool we can remember.

That does not refute the thesis. The authors would say software is the easy case and that the interesting question is whether the same speed appears in medicine or law, where the feedback loops are the ones they describe. We think that is right, and we think the honest reading is that their diffusion lag holds for the domains they chose as examples and does not obviously hold for the domain where most of the current money is being made. A theory that predicts slow adoption everywhere except where adoption is fastest needs to say something about that exception.

What we would want them to add

The essay's closing commitment is to falsifiability. They say the normal-technology view predicts specific observable things, above all adoption rates in consequential domains, and they invite people to check. That is the right posture, and it is more than most essays in this argument offer. We would like to see the predictions written down with dates, so that in 2027 we can score them.

Our own guess is that the essay will look correct about hospitals and courts and wrong about knowledge work that already lives inside a computer. If that split is real, the interesting policy question becomes which parts of the economy get the electrification timeline and which get the software one.