An essay by Sayash Kapoor and Arvind Narayanan, published May 21, 2026 by the Knight First Amendment Institute at Columbia University in its Deep Dive series under the "Toward a Better Internet" channel. It is a direct reply to Derek Thompson, who had engaged with the authors' earlier AI as Normal Technology essay, accepted its account of slow labor-market impacts, and rejected its account of risk (Source: knightcolumbia.org).
Kapoor is a Senior Fellow at Mozilla, a Laurance S. Rockefeller Fellow in the Princeton University Center for Human Values, and a computer science PhD candidate at Princeton's Center for Information Technology Policy. Narayanan is a professor of computer science at Princeton and was the Knight Institute's visiting senior research scientist for 2022–2023.
The disagreement being answered
The authors summarize Thompson's position as accepting the diffusion argument on economics — average GDP growth, unemployment below five percent, and rising employment and wages even in jobs thought vulnerable to automation — while holding that AI's emergent capabilities, unknown even to developers, make it "abnormal" in a way that justifies extraordinary regulation. They quote his framing that treating AI as abnormal "compels the government to create extraordinary regulations that prevent private companies from selling their products and services on the grounds that they're too dangerous."
They note in a footnote that Thompson does not specify what those restrictions would look like in practice, and that their analysis of the necessary interventions is their own reconstruction rather than his proposals.
Definition of extraordinary intervention
Because Thompson does not define the term, the essay supplies a three-factor definition, framed as a spectrum applying to any powerful technology:
- Precautionary — restricting activity based on anticipated rather than realized or demonstrated harms. The authors do not treat precaution as never warranted, but hold that it demands a stronger justification, and that an intervention becomes more extraordinary when less restrictive alternatives exist.
- Burdening actors not responsible for the harm — restrictions on what AI companies may release fall on the builders of dual-use tools rather than on the malicious actors causing harm, and can cut off beneficial access for the public to prevent misuse by a small number of bad actors.
- Bypassing normal governance — relying on unilateral authority such as emergency declarations or executive orders, when the processes being bypassed exist to subject restrictions on liberty to democratic accountability.
An intervention need not satisfy all three; the more that are present, the higher the bar for justifying it.
The argument against nonproliferation
The nuclear comparison is the essay's principal move. Nonproliferation enforcement required the IAEA, the Non-Proliferation Treaty, decades of diplomacy, sustained investment and at times military confrontation, and was workable only because enriched uranium is a genuine physical bottleneck. AI has no equivalent: the core techniques are well known, and adversaries — especially nation-states — can match frontier capabilities within months, so any nonproliferation regime "would face constant erosion."
On chip export controls the authors' assessment is deliberately mild. They call export controls a modest intervention, on the ground that countries routinely restrict exports of sensitive goods, but hold that they are far less effective at preventing dangerous capabilities from becoming widely accessible: open-weight models and broad API access mean "the gap between frontier and publicly available capabilities is at most a few months, not years." See Export Controls (AI), Open-Weight Frontier Models.
The escalation argument follows: to actually enforce nonproliferation, governments would have to restrict open-weight models and API access to capable models through licensing regimes with ongoing authority over releases. The essay cites Anthropic's designation as a supply chain risk and reported discussion of licensing requirements for AI companies as existing steps in that direction (Source: cnbc.com; csis.org). It presses proponents to state where they would draw the line — on open-weight models, per-release approvals, or restrictions on researcher movement across borders — arguing that unstated limits invite escalating claims on unilateral authority as capabilities advance.
A structural point distinguishes the two cases: a nuclear nonproliferation regime, once built, did not need rebuilding every few years, whereas AI is not the last dual-use digital technology, so an "abnormal technology" framework would resemble "a permanent expansion of government powers over what citizens and companies can build, publish, and research."
Three historical episodes are offered as precedent for restraint. The 1995 bill introduced by Senator Feinstein after the Oklahoma City bombing, which in first draft would have criminalized distributing bomb-making information online and passed only in a narrower form requiring knowing aid to a crime. The 1990s attempt to restrict encryption software through export controls and mandated backdoors, including a criminal investigation of a programmer under the Arms Export Control Act, later rolled back through court rulings and executive action, with encryption becoming the foundation of digital security. Against these the essay sets a counter-example — the expansion of FBI surveillance after dynamite-era terrorism — and frames the question as whether AI risks warrant that level of intervention.
The case for resilience
Resilience is defined as the capacity of a system to withstand and adapt to harm, and is presented as imposing no costs on AI companies while improving capacity to respond and recover regardless of where risks arise.
The cybersecurity precedent is developed at length. Internet worms were answered not by restricting access to computers or networks but through bug bounties, improved browsers and operating systems, automated testing and better patching. Automated vulnerability detection is treated as the closer analogy: fuzzers and symbolic execution engines have been superhuman at vulnerability detection for years and are freely available in open-source repositories, and because defenders had the same tools plus structural advantages such as deeper access to the systems under test, they became core defensive tools funded by the cyberdefense ecosystem. The authors accept that the transition was neither smooth nor painless — for a period the attacker-defender balance was upset, and teenagers brought down major e-commerce sites and search engines — and allow that government intervention then could plausibly have reduced harm.
Applied forward, the essay does not predict a smooth AI transition. Under-defended systems named include schools, hospitals, power grids and small government agencies. Anthropic's Project Glasswing and OpenAI's cybersecurity grant program are described as important but insufficient on their own (Source: anthropic.com; openai.com). The proposals are AI-assisted red-teaming extended beyond technology companies to schools, hospitals, power grids, small businesses and government systems, and bug bounties covering more than technology-company products. See AI and Cybersecurity.
The biosecurity section applies the same logic to downstream steps: bioattacks depend on procuring materials, accessing specialized equipment and applying tacit know-how, so resilience can be built now through better screening of synthetic-biology orders, using AI to evaluate the riskiness of new compounds, tracking access to dangerous materials, and offensive red-teaming by trusted experts. These measures help "regardless of whether biorisks are from AI." See AI Biosecurity.
Why resilience has not been prioritized
The essay's answer is about state capacity rather than about AI. Resilience "only requires us to get our act together on the 'normal' process of policymaking and execution," but that process requires polycentric governance in which many decision-makers coordinate, and U.S. state capacity has been hobbled by accumulated veto points and creeping proceduralism. Resilience demands legislation and funding, cross-agency collaboration, early-warning systems and a federal role as information hub — a set of tasks the authors describe the federal government as not well known for performing competently.
Extraordinary action is therefore doubly attractive: morally satisfying, because it burdens the companies creating the risk, and tractable, because it requires only unilateral executive action such as invoking the Defense Production Act. The authors note that even where the "normal" and "abnormal" camps agree — on transparency requirements, auditing, and safe harbors for safety research — no concrete federal action has followed, citing a congressional letter urging better federal coordination on AI-discovered vulnerabilities as a recommendation not yet acted on (Source: latta.house.gov).
The closing argument is about failure modes. If nonproliferation is the main defense, "a single technical breakthrough that makes models cheaper to train could be enough to cause instability," and the damage is greater for the delay in building resilience.
Provenance
Retrieved in full from the Knight First Amendment Institute's own site, with both footnotes and the acknowledgments intact. Two figures — a nonproliferation-versus-resilience illustration and a stylized spectrum of vulnerability-detection capability — are referenced in the raw capture but not reproduced. This is a position essay and should be treated as an argument rather than as evidence.
Relationships
- depends-on: AI as Normal Technology — extends the risk argument of that essay's Part III.
- supports: AI as Normal Technology — the resilience-over-nonproliferation position within it.
- related: Open-Weight Frontier Models — the claim that the frontier-to-open-weight gap is at most a few months.
- related: Export Controls (AI) — characterizes chip export controls as a modest rather than extraordinary intervention.
- related: AI and Cybersecurity, AI Biosecurity, Knight First Amendment Institute (Knight Columbia), Anthropic.