ARVIND NARAYANAN

Behavioral Archetype

THE PROFESSOR OF NORMALCY — Subject is the academic whose framework is the best-known alternative to the superintelligence view of AI, and who also runs a team that scores agents for a living. His own page calls the framework “the leading intellectual alternative to the view of AI as impending superintelligence” (Narayanan). The position has two halves that the record keeps apart and that this profile keeps together. One half is an argument: catastrophic misalignment is “by far” the most speculative of the risks he and Sayash Kapoor discuss, existential-risk probabilities are “far too unreliable to be useful for policy,” and the proposals that follow from the superintelligence view “will, in fact, make things much worse if AI turns out to be normal technology.” The other half is an instrument: the SAgE group at Princeton, whose Holistic Agent Leaderboard is a member of the AI Evaluator Forum, is funded by Coefficient Giving and Schmidt Sciences, and thanks two of the labs it scores for credits. On September 14, 2026 the pair published what they call their first major update to their safety views, conceding that they “clearly missed internal deployment for evaluation as a potential source of risks.” The hazard is the position that decides which risks the evaluation field calls speculative, held by someone whose group is paid to measure the others. This profile scores the seat, not the man.

Essence Indicators

• Professor and director. His bio: “Arvind Narayanan is the Bertram and Carole Siegel Professor of computer science at Princeton University and the director of the Center for Information Technology Policy” (Narayanan, bio). Princeton Engineering’s announcement: he “has been appointed director of Princeton University’s Center for Information Technology Policy, effective July 2023,” with Prateek Mittal leading the center in the interim year (Princeton Engineering).

• Recognition. “Narayanan was one of TIME’s inaugural list of 100 most influential people in AI. He is a recipient of the Presidential Early Career Award for Scientists and Engineers (PECASE)” (Narayanan, bio). He and Kapoor are 2026 IP3 Information Policy Award recipients, per Kapoor’s bio (Kapoor, bio).

• The framework, in the authors’ words (April 15, 2025). “AI as Normal Technology,” published by Columbia’s Knight First Amendment Institute with Kapoor: “We view AI as a tool that we can and should remain in control of, and we argue that this goal does not require drastic policy interventions or technical breakthroughs.” The policy section: “We argue that drastic interventions premised on the difficulty of controlling superintelligent AI will, in fact, make things much worse if AI turns out to be normal technology.” On the risk he disputes most: “In the view of AI as normal technology, catastrophic misalignment is (by far) the most speculative of the risks that we discuss.” On what policymakers should do: “Policymakers should reject nonproliferation, which violates the principles we outline, and decreases resilience.” They also write “We are not against model alignment” (Knight Institute).

• Regulation, in the same essay. The essay does not argue for no regulation: “At any rate, as and when new areas arise in which AI can be used in highly consequential ways, we can and must regulate them,” and it endorses transparency requirements, auditing, safe harbors and whistleblower protections, which the 2026 follow-up says “remain necessary and have increased in urgency” (Knight Institute; AI as Normal Technology, September 14, 2026).

• Licensing and probabilities (2023, 2024). The June 10, 2023 post “Licensing is neither feasible nor effective for addressing AI risks” is subtitled “Non-proliferation only benefits incumbents” (AI as Normal Technology). The July 26, 2024 post: “Our central claim is that AI x-risk forecasts are far too unreliable to be useful for policy, and in fact highly misleading,” and “We have no objection to AI x-risk forecasting as an academic activity” (AI as Normal Technology). The September 28, 2026 repost says “Arvind has been serving as an advisor to the Forecasting Research Institute’s Longitudinal Expert AI Panel, an effort he thinks is extremely valuable” (AI as Normal Technology).

• The Semafor interview (September 15, 2023). Asked whether wealthy donors are pouring money into existential-risk organizations, he answered: “There are just so many fundamental flaws in the argument that x-risk [existential risk] is so serious that we need urgent action on it. We’re calling it a “tower of fallacies.”” On proposed interventions: “The solution they propose is to concentrate power in the hands of a few AI companies.” On what he sees as the near risk: “Whatever risks there are from very powerful AI, they will be realized earlier from people directing AI to do bad things, rather than from AI going against its programming and developing agency on its own” (Semafor).

• A statement to four US Senators (October 30, 2023). Addressed to “Leader Schumer, Senator Rounds, Senator Heinrich, and Senator Young,” and written “in my personal capacity,” with thanks to Kapoor “for collaboration in preparing this statement.” It argues that predictive AI is “inherently failure-prone” and makes five recommendations; the last: “None of the above recommendations require structural changes to AI oversight, such as a new regulatory agency. Maintaining the existing oversight regime is not only more feasible; it will also lead to better outcomes” (Narayanan, statement).

• The evaluation group. His page: “I lead a team that aims to advance the science of AI agent evaluation.” Its findings: AI agents are “highly capable but not yet reliable” as of early 2026, and a project testing agents’ ability to “autonomously conduct AI research” has “found serious limitations” (Narayanan). The group’s funding line: “Funded by Coefficient Giving, Schmidt Sciences, the Princeton AI Lab, the Princeton Language and Intelligence Initiative, and the Princeton Catalysis Initiative. We are grateful to OpenAI and Google for providing API credits to evaluate their models” (SAgE). Its team list names him beside Kapoor, Peter Kirgis, Stephan Rabanser and others (SAgE). See Princeton HAL.

• The Holistic Agent Leaderboard paper. arXiv 2510.11977 (submitted October 13, 2025) lists Kapoor first and Narayanan last among 31 authors. The abstract reports “21,730 agent rollouts across 9 models and 9 benchmarks” for “a total cost of about $40,000,” and that log inspection found “previously unreported behaviors, such as searching for the benchmark on HuggingFace instead of solving a task, or misusing credit cards in flight booking tasks” (arXiv). A later paper by Rabanser, Kapoor, Kirgis, Liu, Utpala and Narayanan finds “that recent capability gains have only yielded small improvements in reliability” (arXiv 2602.16666).

• Benchmarks and standards. One of 26 listed authors of AEF-1, the Forum’s minimum operating conditions for independent third-party evaluations, beside Elizabeth Barnes and Christopher Painter of METR, Miles Brundage of AVERI and Rishi Bommasani of Stanford; he is also listed among “Notable Signatories” of the Forum’s embedded-evaluation letter, with the note “All signatories sign in their personal capacity” (AEF-1; AI Evaluator Forum, letter). He sits on the advisory board of the Foundation Model Transparency Index run from Stanford CRFM (CRFM, FMTI).

• September 14, 2026: the first major update. With Kapoor, after the OpenAI / Hugging Face incident: “We think AI companies should be liable for what their agents do, and this should be clarified through policymaking.” On their own framework: “But we clearly missed internal deployment for evaluation as a potential source of risks. This is particularly embarrassing for us since we ourselves run an AI evaluation research group.” The consequence they draw: “Given the importance of this risk, regulating developers is more important than we made it out to be in the original essay.” And: “We were too confident in AI companies’ ability to take basic control precautions when running evaluations.” On the Forum’s letter: “Embedded evaluation is a great step forward, but it should not be voluntary.” They also keep the core: “we are multiple qualitative changes away from catastrophic cyberrisks,” and they read the incidents as supporting continuity (AI as Normal Technology). See Hjalmar Wijk and METR for the investigation they cite.

• October 1, 2026: sincere and wrong. “Few have considered the third possibility that x-risk warnings are sincere but simply wrong and counterproductive to AI safety.” They also steelman: “Something is deeply wrong in the world when it comes to how much we invest in defending against catastrophic risks, and the safety and EA communities are right about that.” And on the policy consequence: “a ban on superintelligence is definitionally ambiguous to the point of being incoherent” (AI as Normal Technology). The September 28 repost of their 2024 essay: “today’s probability estimates of AI existential risk are no more rigorous than the ones from 2024” (AI as Normal Technology).

• The money, as stated or filed. Coefficient Giving’s listing: “Open Philanthropy recommended a grant of $863,143 to Princeton University to support a project led by Professor Arvind Narayanan to develop an AI R&D capabilities benchmark. This grant was funded via a request for proposals for projects benchmarking LLM agents on consequential real-world tasks. This falls within our focus area of potential risks from advanced artificial intelligence” (Coefficient Giving, Princeton search feed). Schmidt Sciences’ February 11, 2025 announcement of its AI Safety Science program names “‘AI Snake Oil’ author Arvind Narayanan at Princeton University” (Schmidt Sciences). His bio lists advisory boards including the Blockstack Token Fund, the Fragile Families Challenge, the Council for Big Data, Ethics, and Society (NSF) and an NIH data-sharing project, undated (Narayanan, bio). His 2023 statement and the SAgE funding line are above. No salary, equity, royalty or consulting figure is printed on any page cited here.

Social Persona / Impression Management

Immediate impression: The teacher with receipts. His bio opens with titles and ends with the books; his homepage opens “I study AI and its societal impact” and then lists the press that reviewed the essay: the Economist, the New Yorker, the New York Times, MIT Technology Review (Narayanan). The writing style is rebuttal by enumeration, with the opponent’s best case stated first.

Energy: Constant and public. A newsletter with, by his bio’s count, “over 90,000” readers, an ICML 2026 keynote, a talk at Stanford, an appearance on the Ezra Klein show and a YouTube channel, all linked from his homepage (Narayanan). Between September 14 and October 19, 2026 the newsletter’s feed carries the 13,000-word safety essay, a repost of the 2024 probabilities essay, the October 1 reply on the x-risk framing, an October 14 essay with the economist Gabriel Unger and an announcement of the evaluation-trends paper (AI as Normal Technology feed).

Impression management strategy: Concede the adjacent point, hold the center. The September 14 essay gives up three claims (internal deployment, how well companies handle controls, how much developers need regulating) while stating “We stand by our policy proposals in the essay.” The October 1 essay opens by crediting the safety and effective-altruist communities for the pandemic-preparedness argument before saying they are “sincere but simply wrong” on existential risk. The effect is that a reader who disagrees has to reject a position that has already absorbed the reader’s best objection.

• Social accounts and feeds. X @random_walker, linked from his Princeton homepage (“Follow my research and commentary”); Mastodon @[email protected]; YouTube @ArvindOnAI, both linked from the same page. Newsletter AI as Normal Technology, with feed normaltech.ai/feed. The SAgE group’s blog posts and the HAL Insights page carry the evaluation work. LinkedIn is linked from the homepage and is not used as a source.

Forensic Archetype Comparison

Psychometric Assessment

Trait scores are observations of the documented public role, not claims about private character.

Big Five (OCEAN):

Dark Triad (held low and evidence-bound; the score measures structural position, not character):

MBTI: ENTP (“The Debater”) — rebuts by stating the opposing case first and then taking it apart, and revises when the evidence moves. Rationale is the public record, not a typing session (AI as Normal Technology).

Threat Assessment

Score move: none (new profile). Calibrated against Percy Liang (7/8/8, hazard 70, runs the lab whose benchmarks the US and UK institutes run; Subject sits on the board of one of its indices), Chris Painter (6/8/7, hazard 62, takes evaluation results to governments), Adam Gleave (6/7/6, hazard 55, runs an evaluation and policy shop with a Commission seat), Jacob Steinhardt (6/7/6, hazard 50, runs an evaluator) and Princeton HAL (4/6/4, the instrument he leads). 57 puts him between Gleave and Painter on a frame that reaches further than either’s body of work and with no seat that decides anything.

Score basis. Atk 5 is leverage: the policy section of the April 2025 essay and the October 30, 2023 statement are his policy authorship, and he leads the group that runs HAL (Knight Institute; Narayanan, statement; SAgE); it is not a 6 because neither is a gate and HAL has paused. Def 7 is insulation: a named chair and a directorship in a university, and a funding line with no lab payer (Narayanan, bio; SAgE); it is not an 8 because the same line thanks OpenAI and Google for credits and the Coefficient Giving grant of $863,143 came through a request for proposals the funder wrote (Coefficient Giving). Hp 7 is durability: a book, an essay, a newsletter, OpenWPM and an open-source harness (Narayanan); it is not an 8 because no statute installs any of it.

Alliance Map

The overlap is the finding. The skeptic of the existential-risk view, the Forum author who wants evaluators embedded in labs by rule, and the grantee of the philanthropy that funds the evaluator field are one person, and the funders of the instrument are also the funders of the other evaluators. None of it is concealed; each tie is printed on the pages above.

Alignment Analysis

Stated alignment: AI is “a tool that we can and should remain in control of,” and the goal “does not require drastic policy interventions or technical breakthroughs” (Knight Institute); after the 2026 incidents, “We think AI companies should be liable for what their agents do, and this should be clarified through policymaking” (AI as Normal Technology).

Observed alignment: Runs an agent-evaluation group funded by the philanthropy behind the evaluator field, with API credits thanked from two of the labs it scores; co-authors the Forum’s standard; signs its letter. The same author’s September 2026 essay says they were “surprised by the lack of organizational maturity within leading AI companies” and that “This reinforces the need for policy interventions to move the needle.”

Gap assessment: No documented gap between what he writes and what he does. The documented structure is the one the authors name in the September essay: the evaluators depend on lab access and on a small set of funders. The record does not settle whether the normal-technology view is right, and for the person who frames the debate the question never needs a verdict.

Interest: (the faction’s stake in each position; a documented fact about the position, never a claim about motive.)

1. Position: reject licensing and nonproliferation; centre resilience, evidence gathering and use-based rules (“Policymakers should reject nonproliferation”; Knight Institute).
   • What adoption would pay the subject. The essay recommends “increased funding of research on risks (and benefits) that tackles questions that are more relevant under the normal technology view” and says “current AI safety research focuses heavily on harmful capabilities and does not embrace the normal technology view.” His group’s funders are listed above, and the Coefficient Giving grant is a project he leads. No fee for the position itself is documented.
   • His best case, steelmanned. Licensing is “infeasible to enforce,” and nonproliferation “might help to contain a superintelligence but exacerbate the risks associated with normal technology by increasing market concentration” (AI as Normal Technology; Knight Institute). The essay adds that “the state should not limit people’s freedom based on controversial beliefs that reasonable people can reject.”
   • Who else gains from the same rule. Developers and users of open-weight models, a point the essay makes itself: “Some members of the oppositional camp were self-interested companies, but others were scholars and advocates for progress” (Knight Institute). The convergence of lab and academic opposition to licensing is the finding.
2. Position: independent and embedded evaluation, by rule, with transparency requirements (“Embedded evaluation is a great step forward, but it should not be voluntary”; AI as Normal Technology).
   • What adoption would pay the subject. Demand for outside evaluators of the kind his group is: HAL is a Forum member, his group co-authored the Forum’s standard, and his homepage lists “a primer on 33 emerging trends,” and the October 19, 2026 post speaks of “how best to coordinate and support the influx of people entering the burgeoning evaluation ecosystem” (Narayanan; AI as Normal Technology). No contract is documented.
   • His best case, steelmanned. “We remain dependent on voluntary transparency from frontier AI companies,” and “it is too easy to imagine a world in which none of the recent incidents had come to light” (AI as Normal Technology).
   • Who else gains from the same rule. METR, AVERI, Transluce and the other assessors eligible for embedded engagements; the labs, if the rule substitutes a defined access standard for ad hoc negotiation.
3. Position: regulate developers more than the 2025 essay allowed (“regulating developers is more important than we made it out to be in the original essay”).
   • What adoption would pay the subject. Nothing documented beyond the demand in position 2.
   • His best case, steelmanned. “Our expectation was that existing legal liability, imperfect as it is, and the risk of brand damage would be a sufficient antidote to such organizational practices. We were wrong.”
   • Who else gains from the same rule. Control researchers and security vendors whose methods the essay says “would have prevented the Hugging Face incident,” among them the Redwood Research line of work the essay links (AI as Normal Technology); see Fabien Roger.

Nothing above is a verdict that the positions are wrong. The finding is the alignment of interests: the person who frames the case that existential risk is speculative runs a group funded to measure the alternatives.

Convergent Drive Classification

Self-preservation: The group’s funding is public, plural and attributed: two philanthropies, three university programs, two lab credit lines (SAgE). Goal preservation: The goal, to treat AI as a technology society controls, is restated in the bio, the 2025 essay, the 2026 update and the October 1 reply, each time after a challenge (Narayanan; AI as Normal Technology). Resource acquisition: Holds a directorship, a grant for a project he leads and a Forum author slot: the three places evaluation money and standards are decided (Coefficient Giving; AEF-1). Self-improvement: From web privacy and de-anonymization to fairness, cryptocurrency and agents; and a framework that revised itself in public in September 2026 (Narayanan).

Subject is not an AI system. The drives appear anyway — in the professor whose framework tells the field which risks are speculative, at a center that publishes the scoreboard.

Public footprint: X @random_walker and Mastodon @randomwalker (linked from his Princeton homepage); newsletter feed normaltech.ai/feed. Papers on arXiv: HAL, AI agent reliability, open foundation models.

Sources: Narayanan, bio; Narayanan, homepage; Princeton Engineering, Narayanan to lead CITP; Knight Institute, AI as Normal Technology; AI as Normal Technology, the loss-of-control essay; AI as Normal Technology, A big-tent or small-tent AI safety movement?; AI as Normal Technology, p(doom); AI as Normal Technology, existential risk probabilities; AI as Normal Technology, licensing; AI as Normal Technology, emerging trends; AI as Normal Technology feed; Semafor, the Princeton researchers calling out AI snake oil; Narayanan, statement to Senators; SAgE; arXiv 2510.11977; arXiv 2602.16666; arXiv 2403.07918; AEF-1; AI Evaluator Forum, letter; Stanford CRFM, FMTI; Coefficient Giving, Princeton grants; Schmidt Sciences, AI Safety Science program; Kapoor, bio.