Arvind Narayanan and Sayash Kapoor · May 21, 2026
Sayash Kapoor and Arvind Narayanan argue in Normal Technology for resilience
Markdown · JSON · Original source
Sayash Kapoor and Arvind Narayanan published this essay on May 21, 2026. They argue that addressing AI misuse risks requires investment in resilience—the capacity of systems to withstand and adapt to harm—rather than relying on extraordinary restrictions on what companies can release. They say AI nonproliferation is brittle: AI lacks the physical bottleneck that makes nuclear nonproliferation more enforceable, and adversaries can match frontier capabilities within months. Restrictions on access, they warn, may buy only a few months while expanding government authority over what people can build, publish, and research. The authors’ alternative is to strengthen defenses across society. They point to cybersecurity practices and call for AI-assisted and broader bug bounties for under-defended institutions, alongside downstream measures to improve biosecurity. These efforts require investment and planning, they say, and government resilience work remains inadequate. They understand why unilateral restrictions may seem more tractable, but toward improving ordinary policymaking and execution. They end by warning that relying on nonproliferation without resilience could mean greater damage if a technical breakthrough makes models cheaper to train.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
Arvind Narayanan and Sayash Kapoor · Article
Source details
Original source date: May 21, 2026 (2026-05-21T13:19:41.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-07T03:03:26.079Z. First observed: 2026-10-23T20:39:43.075Z. Retrieved: 2026-10-23T20:41:38.099Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 23, 2026; originally published May 21, 2026. The essay describes Derek Thompson’s position and proposals as vague; his essay is not included in the supplied evidence. The authors identify proposed resilience measures but do not specify the scale of investment or how effective those measures would be.
Open question
What resilience investments and government actions would be needed to make the authors’ proposed approach effective across the systems they identify?
Source-stated
The authors argue that resilience is a defense against AI misuse risks that does not require restricting what AI companies can release.
resilience does not impose costs on AI companies. Instead, it focuses on improving our capacity to respond to and recover from AI risks
Source-stated
The authors argue AI nonproliferation is less enforceable than nuclear nonproliferation because AI lacks a physical bottleneck and adversaries can match frontier capabilities within months.
there is no equivalent “physical” bottleneck. The core techniques for building AI systems are well known. Adversaries (especially nation-states) can match frontier capabilities within months.
Source-stated
The authors say open-weight models and API access leave a gap of at most a few months between frontier and publicly available capabilities.
Open-weight models and widespread API access from frontier labs mean that the gap between frontier and publicly available capabilities is at most a few months, not years.
Source-stated
The authors argue resilience requires investment and planning, including AI-assisted red-teaming and vulnerability reporting beyond technology companies.
That means AI-assisted red-teaming not just for tech companies, but for schools, hospitals, power grids, small businesses, and government systems that currently lack the capacity for defense.
Source-stated
The authors lean toward improving ordinary policymaking and execution, while warning that relying on nonproliferation without resilience could increase damage if a breakthrough makes models cheaper to train.
Getting our policy act together is hard, but important—not just for this round of AI policy and misuse risks, but for all future interventions on technological harm, and for the democratic process to work more generally.
Source and contributor context
Arvind Narayanan and Sayash Kapoor · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Red-teaming
Deliberately attacking or tricking a system to find its weaknesses before others do.
Lean
Lean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
Editorial provenance
Version 1. Publication ID: 1226a889-cbd6-ca6a-41a3-37aebfda298b.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T02:59:26.997Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: c2b1060bb9991f697eaa32b383a09dce857c37b09a646232bb90da999071b33f. Source text hash: ed601c67eb67710cdfd56cbc0ca2871fc624aed1e46d2ced6d9bec3d491e5cc7.
Revision and source-check history
Version 1 · 2026-11-07T03:03:26.079Z
Sayash Kapoor and Arvind Narayanan argue in Normal Technology for resilience
Sayash Kapoor and Arvind Narayanan published this essay on May 21, 2026. They argue that addressing AI misuse risks requires investment in resilience—the capacity of systems to withstand and adapt to harm—rather than relying on extraordinary restrictions on what companies can release. They say AI nonproliferation is brittle: AI lacks the physical bottleneck that makes nuclear nonproliferation more enforceable, and adversaries can match frontier capabilities within months. Restrictions on access, they warn, may buy only a few months while expanding government authority over what people can build, publish, and research. The authors’ alternative is to strengthen defenses across society. They point to cybersecurity practices and call for AI-assisted and broader bug bounties for under-defended institutions, alongside downstream measures to improve biosecurity. These efforts require investment and planning, they say, and government resilience work remains inadequate. They understand why unilateral restrictions may seem more tractable, but toward improving ordinary policymaking and execution. They end by warning that relying on nonproliferation without resilience could mean greater damage if a technical breakthrough makes models cheaper to train.
Original source
Arvind Narayanan and Sayash Kapoor · Article
Source details
Original source date: May 21, 2026 (2026-05-21T13:19:41.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-07T03:03:26.079Z. First observed: 2026-10-23T20:39:43.075Z. Retrieved: 2026-10-23T20:41:38.099Z.
Limits
Text as retrieved on October 23, 2026; originally published May 21, 2026. The essay describes Derek Thompson’s position and proposals as vague; his essay is not included in the supplied evidence. The authors identify proposed resilience measures but do not specify the scale of investment or how effective those measures would be.
Open question
What resilience investments and government actions would be needed to make the authors’ proposed approach effective across the systems they identify?
Source-stated
The authors argue that resilience is a defense against AI misuse risks that does not require restricting what AI companies can release.
resilience does not impose costs on AI companies. Instead, it focuses on improving our capacity to respond to and recover from AI risks
Source-stated
The authors argue AI nonproliferation is less enforceable than nuclear nonproliferation because AI lacks a physical bottleneck and adversaries can match frontier capabilities within months.
there is no equivalent “physical” bottleneck. The core techniques for building AI systems are well known. Adversaries (especially nation-states) can match frontier capabilities within months.
Source-stated
The authors say open-weight models and API access leave a gap of at most a few months between frontier and publicly available capabilities.
Open-weight models and widespread API access from frontier labs mean that the gap between frontier and publicly available capabilities is at most a few months, not years.
Source-stated
The authors argue resilience requires investment and planning, including AI-assisted red-teaming and vulnerability reporting beyond technology companies.
That means AI-assisted red-teaming not just for tech companies, but for schools, hospitals, power grids, small businesses, and government systems that currently lack the capacity for defense.
Source-stated
The authors lean toward improving ordinary policymaking and execution, while warning that relying on nonproliferation without resilience could increase damage if a breakthrough makes models cheaper to train.
Getting our policy act together is hard, but important—not just for this round of AI policy and misuse risks, but for all future interventions on technological harm, and for the democratic process to work more generally.
Source and contributor context
Arvind Narayanan and Sayash Kapoor · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Red-teaming
Deliberately attacking or tricking a system to find its weaknesses before others do.
Lean
Lean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
Editorial provenance
Version 1. Publication ID: 1226a889-cbd6-ca6a-41a3-37aebfda298b.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T02:59:26.997Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: c2b1060bb9991f697eaa32b383a09dce857c37b09a646232bb90da999071b33f. Source text hash: ed601c67eb67710cdfd56cbc0ca2871fc624aed1e46d2ced6d9bec3d491e5cc7.
uncertain · 2026-11-07T02:59:26.997Z · anthropic/claude-opus-5.5
Publication: 1226a889-cbd6-ca6a-41a3-37aebfda298b
attribution — pass: The essay is correctly attributed to Sayash Kapoor and Arvind Narayanan on their Normal Technology publication. Its positions are framed as the authors' arguments, not as established fact.
Sayash Kapoor and Arvind Narayanan May 21, 2026
summary — uncertain: The summary is faithful, but two gaps limit a general reader's understanding. First, the central term 'nonproliferation' is never explained; here it means keeping advanced AI capabilities out of wide availability. 'Open-weight' and 'frontier capabilities' are also unexplained, and only red-teaming is defined. Second, it omits that the essay answers Derek Thompson's case for 'extraordinary' government action, which frames the authors' argument. Smallest repair: one clause defining nonproliferation and one naming Thompson's thesis as the position being rebutted.
In this essay, we lay out the downsides of extraordinary government intervention in response to new technology.
date — pass: May 21, 2026 matches originalPublished, the essay's dateline and modifiedAt. The retrieval date is disclosed. Nothing in the reading uses later events or hindsight. The kind is publisher-page, so no archive disclosure is needed.
May 21, 2026
support — pass: Each claim is supported by its excerpt and attributed to the authors. This includes the few-month gap, the missing physical bottleneck, resilience measures, the lean toward normal policymaking and the closing warning about a training-cost breakthrough. The question about needed investments follows from the essay without presuming an answer, and the limitations are accurate.
a single technical breakthrough that makes models cheaper to train could be enough to cause instability, especially in a world where we don’t also invest in resilience
media — pass: No media was selected, and no verified candidates were supplied, so there is no media identity or relevance issue.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Deliberately attacking or tricking a system to find its weaknesses before others do.
Definition referenceLean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
Explore LeanDeliberately attacking or tricking a system to find its weaknesses before others do.
Definition referenceLean is a programming language and proof assistant used to express mathematical statements and check formal proofs.
Explore Lean