Princeton professor and CITP director, AI policy and evaluation researcher
Arvind Narayanan
Profile
Arvind Narayanan is what the AI field badly needs more of: a serious computer scientist who reads the papers, runs the experiments, and then tells you plainly when the emperor has no clothes. He’s a professor of computer science at Princeton and director of its Center for Information Technology Policy, and for developers trying to separate genuine capability from marketing, his work is close to essential reading. His trajectory matters here: this isn’t a pundit lobbing takes from the sidelines. Narayanan earned his PhD at UT Austin under Vitaly Shmatikov, did a postdoc at Stanford with Dan Boneh, and built a research career on privacy and de-anonymization before AI hype became his beat. As a grad student he and Shmatikov showed that the “anonymized” Netflix Prize dataset could be re-identified by cross-referencing public IMDb reviews — a result that permanently changed how people think about big-data privacy. He later led the Princeton Web Transparency and Accountability Project measuring online tracking, and co-authored the widely used free textbook Bitcoin and Cryptocurrency Technologies.
Today he’s best known as co-author, with Sayash Kapoor, of AI Snake Oil (2024) and the newsletter of the same name, now published under the AI as Normal Technology banner and read by tens of thousands of researchers, policymakers, and journalists. The through-line of all of it is empirical skepticism: not the reflexive AI-is-useless contrarianism of some critics, and not the breathless superintelligence-is-imminent framing of the labs, but a demand that claims be backed by evidence. Narayanan is happy to say that large language models are genuinely useful and transformative — and in the same breath dismantle a benchmark result or a “predictive AI” product that doesn’t survive scrutiny.
That evidence-first stance is what makes him useful to people actually building with AI. His critique of “GPT-4 passed the bar exam at the 90th percentile” — data contamination, misleading percentile framing, and the gap between test-taking and real professional work — is the kind of analysis every developer citing a leaderboard should internalize. With Kapoor he has pushed hard on the evaluation crisis in agents, showing that state-of-the-art agents are often needlessly complex and expensive, and that ignoring cost while chasing accuracy produces misleading rankings. His broader “Leakage and the reproducibility crisis in ML-based science” work documented how data leakage has corrupted results across hundreds of papers in seventeen scientific fields — a sobering reminder for anyone shipping an ML model.
His current intellectual project, “AI as Normal Technology,” is his most ambitious argument yet: that AI will diffuse into society the way electricity and the internet did — slowly, unevenly, gated by institutions and adoption — rather than arriving as a sudden superintelligent break. It has become the leading counter-framework to the AGI-is-imminent worldview, and whether or not you buy it in full, it’s a more rigorous way to reason about deployment, timelines, and policy than most of what circulates online. Narayanan appeared on the inaugural TIME 100 AI list, is a faculty associate at Harvard’s Berkman Klein Center, and remains one of the clearest thinkers on where AI actually delivers and where it’s being oversold.
Books
Key Articles & Papers
AI as Normal Technology AI Agents That Matter GPT-4 and Professional Benchmarks: The Wrong Answer to the Wrong Question Leakage and the Reproducibility Crisis in Machine-Learning-Based Science Pitfalls of Evidence-Based AI Policy How to Break Anonymity of the Netflix Prize DatasetVideos
Spotify Podcasts
YouTube
