AI Snake Oil: What Artificial Intelligence Can Do, What It Can't, and How to Tell the Difference is a 2024 book by Arvind Narayanan and Sayash Kapoor (Princeton), published in September 2024 by Princeton University Press. It argues that the single label "AI" is applied to several distinct technologies with very different track records, and offers heuristics for distinguishing genuine capabilities from vendor marketing.
Authors: Arvind Narayanan and Sayash Kapoor (Princeton). Published: September 2024 (Princeton University Press). Subtitle: What Artificial Intelligence Can Do, What It Can't, and How to Tell the Difference.
Summary of argument
The book separates three technologies that it says are frequently conflated under "AI." The first is predictive AI, used by institutions to forecast the future, such as criminal recidivism, loan default, and job performance. The authors characterize predictive AI as largely snake oil, arguing it is limited by the irreducible difficulty of predicting human behavior: vendors repackage statistical tools and marketing as prediction, institutions adopt them under procurement pressure, and error rates are often little better than random.
The second is generative AI, which the authors describe as genuinely powerful for some tasks but overhyped. They distinguish real capabilities such as language generation, coding assistance, and summarization from what they call fictional ones such as autonomous agency, reliable factuality, and superhuman reasoning, and argue that a capability-deployment gap is real.
The third is content moderation AI, which the authors say is hard in principle. They argue that classifying speech cannot be solved by better models alone, because context-dependence and shifting norms mean the problem has no stable solution.
The authors argue that AGI hype inflates all three by promising that the problems of predictive AI and content moderation will eventually be solved by scaling generative AI.
Key claims
On predictive AI, the authors marshal chapter evidence on COMPAS (recidivism), hiring tools (HireVue, Pymetrics), civil-service benefits-fraud detection, and academic-dropout prediction, arguing that error rates are much higher than marketers claim and that there is very little accountability when predictions harm individuals. They argue that the bias debate is a distraction when a tool is fundamentally unreliable, on the grounds that debiasing a tool only slightly better than random is not useful.
The authors describe a hype-cycle dynamic in which early capability demonstrations, amplified by the press, produce institutional adoption before the limitations emerge at scale.
On elections and synthetic media, the book argues that deepfakes have not destroyed elections, a position the authors present as consistent with Persily's Digitalist Papers essay: the "fake news decides elections" narrative is empirically weaker than claimed, and the real democracy cost is the liar's dividend.
The authors connect these claims to a set of policy positions. They argue that regulatory focus should fall on accountability at the point of deployment rather than on model-level regulation, and call for AI procurement reform under which agencies would demand evaluation data with meaningful baselines. They also call for press literacy and a refusal to accept vendor metrics at face value, and for decision rights and due-process protections when AI tools affect individuals.
Relation to the "Normal Technology" framing
The book is the popular-press articulation of an argument the authors develop more formally in AI as Normal Technology, a companion academic paper (Knight Columbia, 2025). Both argue that AI is a general-purpose technology with decades-long diffusion rather than a superintelligence-imminent event.
Relationships
- supports: AI as Normal Technology — same authors, extended theoretical framing.
- related: The Digitalist Papers (Stanford, Volumes 1–2) — Persily's democracy-panic essay makes the same point empirically for deepfakes; Cochrane overlaps on regulation-as-capture critique.
- contradicts: AI 2027 / Situational Awareness: The Decade Ahead — explicitly argues against the short-timeline superintelligence framing.
- related: (Source: hai.stanford.edu) / MIT NANDA — The GenAI Divide (State of AI in Business 2025) — empirical evidence that corporate AI deployments underperform claims.
- related: Statement on AI Risk (CAIS) / FLI — Pause Giant AI Experiments: An Open Letter — the existential-risk framing the book explicitly pushes back against.
- related: Jagged Frontier — "jagged frontier" capability discourse that the book critiques as AGI hype.
- related: FTC 6(b) Staff Report: Partnerships Between Cloud Service Providers and AI Developers — regulatory pressure points the book's agenda aligns with.