Arvind Narayanan

In the [YOU] on AI Field Guide

The cycle's project is to see the machine clearly—without the narcotic of hype or the paralysis of fear—and Narayanan is the cycle's primary methodologist for this act of seeing. His contribution is not a philosophical framework or a historical analogy but a set of habits for testing claims that remain useful as the technology evolves: demand evidence proportioned to the claim, place the burden on the party that profits from belief, refuse to be moved by fluency, and redirect attention from the spectacular to the actual. The cycle applies these habits throughout its analysis of AI's effects on professional practice, creative work, and institutional decision-making—and particularly in its treatment of the tools that claim to predict human outcomes, which the cycle distinguishes sharply from the generative tools it advocates.

Narayanan's most structurally important contribution to the cycle is his account of broken institutions and the tools they adopt. The observation that broken AI is appealing to broken institutions—that the flawed automation does not succeed despite its flaws but because of what the institution is trying to avoid—reframes the entire conversation about AI governance. The problem is not that institutions are foolish; it is that they are broken in ways the technology lets them avoid confronting, and that the avoidance is the most expensive thing of all. Fixing the harms that predictive AI produces requires fixing institutions, not buying better software, and the cycle's recommendations for organizational adaptation draw directly on this insight.

The “normal technology” thesis—Narayanan's most ambitious reframing, developed in a 2025 essay with Kapoor—provides the cycle with its most important counterweight to both the rapturous and the apocalyptic accounts of what AI represents. If AI is a nascent species racing toward superintelligence, the central problem is alignment and the response may be drastic. If AI is normal technology—a powerful tool in the lineage of electricity and the internet, diffusing through society at the pace of society's capacity to absorb it—then it is something we can and should remain in control of, and the response is the ordinary, unglamorous work of governance rather than emergency measures premised on imminent loss of control.

Origin

Born in India and trained at the Indian Institute of Technology Madras, Narayanan took his doctorate at the University of Texas at Austin, did postdoctoral work at Stanford, and joined the faculty at Princeton in 2012, where he now directs the Center for Information Technology Policy. The biography matters because it establishes the credential that counts in his kind of argument: he has done the work he is evaluating. His doctoral research on the re-identification of supposedly anonymous datasets produced results of immediate practical consequence—demonstrating that the assurances of privacy through anonymization that companies routinely gave were, in high-dimensional datasets, mathematically unfounded—and established the intellectual posture that defines his career: a comfortable consensus examined by someone willing to do the technical labor of checking, and the consensus failing the check.

The work on anonymization and the work on AI hype are not two separate careers but one continuous investigation. The data economy that his early research exposed—vast accumulations of personal information whose supposed protections he had shown to be illusory—is the same economy that fuels the predictive AI systems he now criticizes. Both involve institutions making confident claims about what their systems can and cannot do, both involve the public lacking the literacy to evaluate those claims, and both demand the same remedy: evidence, transparency, and the right to understand and contest the systems that make consequential decisions about one's life.

AI Snake Oil crystallized two decades of this work into its most accessible and consequential form. Named to the inaugural TIME100 list of the most influential people in AI alongside Kapoor, the book achieved something rare in technology discourse: it gave the public a vocabulary precise enough to make the relevant distinction without requiring technical training. The nineteenth-century snake oil peddler did not sell empty water; he sold something real that produced a real sensation and came wrapped in the language of medicine, and the genius of the fraud was that you could not easily tell the tonic from the poison without knowing something the peddler was counting on you not knowing. Narayanan's claim is that contemporary predictive AI occupies the same position, and his project is to install the literacy that makes the distinction visible.

Key Ideas

The generative-predictive divide. Narayanan's central distinction separates two fundamentally different technologies that share a name and almost nothing else. Generative AI produces plausible continuations of patterns in training data; its failures are visible and correctable. Predictive AI claims to forecast what a particular person will do next; its failures are hidden behind the irreducible uncertainty of the future it purports to describe. The divide is not merely technical—it determines the kind of harm each technology produces, the kind of governance each requires, and the kind of evidence that can demonstrate either one's fitness for a consequential role. Collapsing the two under a single word allows the genuine achievements of generative AI to lend their credibility to the far more dubious claims of predictive AI.

The ceiling on life prediction. The most rigorous evidence in Narayanan's account comes from a large collaborative study in which hundreds of researchers were given an exceptionally rich dataset—thousands of families followed over many years—and challenged to predict outcomes. The result was humbling: the best models predicted life outcomes only slightly better than a simple formula using a handful of variables. The lesson is structural rather than contingent: human lives are shaped by contingency, by events that have not yet happened, by the free responses of the person being predicted to the very prediction being made. No quantity of historical data can encode a genuinely open future, and the difficulty of predicting life outcomes is not a temporary limitation of current models but a feature of the world that no AI architecture will overcome.

Broken AI for broken institutions. The most penetrating insight in Narayanan's account is institutional rather than technical. Flawed predictive systems persist not despite their inadequacy but because of what the institution is trying to avoid: a hiring process drowning in applications cannot read every resume, and the algorithm makes the unmanageable feel manageable. The prediction’s accuracy is almost beside the point; its function is to absorb a volume that the institution cannot otherwise face. The constructive implication is that fixing predictive AI harms requires fixing institutions—asking why the hiring process generates a thousand applications it cannot read, why the healthcare system rations care by formula, why the court needs an algorithm to tell it whom to fear. These are hard, political, expensive questions, and predictive AI is appealing precisely because it promises to make them disappear.

AI as normal technology. Narayanan's most ambitious reframing proposes that AI belongs in the same category as electricity, the printing press, and the internet—transformative general-purpose technologies that reshaped civilization while remaining, in the relevant sense, tools humans built, deployed, governed, and absorbed over time. The normal technology frame targets the dominant alternative, which treats AI as a nascent species racing toward superintelligence. The frame is not complacency: it insists that the harms Narayanan has documented are real and require serious response. But they are the harms of a powerful tool misused or misunderstood, not the harms of an alien intelligence, and the distinction matters enormously for what we do about them.

The anatomy of hype. Narayanan's sociology of AI hype identifies three mutually reinforcing sources: companies with every incentive to overstate capabilities and no incentive to disclose limitations; journalists operating under incentives that reward the dramatic over the accurate; and researchers whose failures of scientific rigor—chief among them data leakage and improper evaluation—generate inflated results that the press and companies broadcast. The benchmark score is the engine of the field's sense of progress, and Narayanan insists it is a narrow and easily corrupted indicator: what a system achieves on a curated test is not what it achieves in deployment, and the gap between demonstration and product is where most of the disappointment lives.