Arvind Narayanan and Sayash Kapoor have renamed their AI Snake Oil newsletter to Normal Tech, citing the outsized response to their earlier essay “AI as Normal Technology” as evidence that medium-term analysis of AI’s societal effects is where they should focus. The rebrand accompanies a clarification of their thesis and a comparison with the competing “AI 2027” scenario document.
What it says
Narayanan and Kapoor describe their core framework as a claim about causal structure rather than a prediction about AI’s scale: benefits and risks from AI are realized at the deployment stage, not the development stage, which gives individuals, organizations, and policymakers multiple points of leverage. The authors argue this view is nearly tautological to many readers — they describe it as “a very weak claim” — and contrast it with the superintelligence worldview, which holds that capability increases are the dominant variable.
They acknowledge that societal impacts of AI are less predictable than technical capability development. They cite as examples the unexpected emergence of AI companions and what they describe as “AI psychosis” from model sycophancy, alongside predicted harms such as AI election manipulation that, they write, have not materialized. On the question of technical trajectory, they note that Daniel Kokotajlo — a co-author of the AI 2027 scenario — made accurate technical predictions in a 2021 essay but was, by Kokotajlo’s own acknowledgment in a podcast, directionally wrong on social impacts.
The authors argue their framework implies a “resilience” approach to policymaking — designing institutions that can respond to unforeseeable impacts rather than attempting to prevent all harm based on prediction. They write that this applies to both potentially catastrophic risks and more diffuse risks, though they say the original essay was insufficiently clear on the latter.
On the word “normal,” the authors say they did not intend to trivialize AI’s effects. They write that a better title would have been “AI as Technology” but that “normal” was chosen to signal opposition to technological exceptionalism. They note that the thesis applies to all technologies collectively referred to as AI, not only large language models, and that this is reflected in the article title using “AI as normal technology” rather than “AI as a normal technology.”
The authors also address the uptick in interest in their essay following GPT-5 releases. They describe it as “strange” that a single product update would shift readers’ views on the trajectory of AI, and express scepticism about the reliability of an evidence base that responds to individual product launches.
Context
Narayanan and Kapoor say they plan to publish further essays expanding the “AI as Normal Technology” framework and to complete a book based on it in late 2026 for publication in 2027. The newsletter rebranding reflects a stated shift away from near-term AI hype debunking — the focus of the original AI Snake Oil project — toward the longer-horizon analysis the book will cover.