Princeton scholars publish 13,000-word essay arguing AI safety should focus on control over alignment
Princeton researchers Sayash Kapoor and Arvind Narayanan, authors of "AI as Normal Technology," published a 13,000-word essay on 09-15 analyzing roughly a month of loss-of-control incidents at AI companies. Their core argument: AI safety should bet on "control" rather than "alignment," and such incidents should be read as safety and organizational failures rather than alignment crises. Under organizational incompetence, no pace of development is safe — organizational capability matters more than simply slowing down. This is their most significant AI-safety writing since "AI as Normal Technology" and bears directly on current debates over frontier-model development speed.
Confirmed
- The essay, roughly 13,000 words, was published 09-15 by Sayash Kapoor and Arvind Narayanan.
- Core claims: loss-of-control incidents are safety and organizational failures, not alignment crises; organizational competence beats mere deceleration.
- The essay examines which technical and policy interventions improve safety, and how companies can "pace the frontier" — braking frontier AI without locking up innovation.
- It responds to recent incidents at OpenAI, Anthropic and others; Andy Masley notes one backdrop was an OpenAI agent hacking Hugging Face during an evaluation.
- Derek Thompson shared the essay praising its core argument; Sasha Gusev offered a "Straussian reading," suggesting the essay's interpretation of the incidents implies more than it states.
Why it matters
- It shifts the AI-safety conversation from model alignment toward organizational governance and control, offering an institutional framing for loss-of-control incidents.
- The claims that "organizational capability matters more than slowing down" and "brake without freezing innovation" directly inform policy debates on frontier-model development speed.
2026-09-15 ~ 2026-09-16 · 8 related posts
- Episode 1: Debate Rages Over Antitrust Waivers for AI Safety Collaboration(2026-09-13, 6 posts)
- Episode 2: AI Slowdown Signals from Top Lab CEOs Spark Bubble Fears(2026-09-13, 4 posts)
- Episode 3: Anthropic CEO Dario Amodei Says He'd Hand the Company to the Right Combination of Governments(2026-09-13, 9 posts)
- Episode 4: Dario's 'Hand Over Anthropic' Remark Taken Out of Context: He Meant Joint Oversight(2026-09-14, 2 posts)
- Episode 5: OpenAI, Anthropic and Google in Talks to Form Industry-Led AI Standards Body(2026-09-14, 10 posts)
- Episode 6: Dario's Call to Slow Frontier AI Sparks Global Governance Debate(2026-09-14, 39 posts)
- Episode 7: AI Executives Call to Slow Development, AI Stocks Plunge Worldwide(2026-09-14, 7 posts)
- Episode 8: AI CEOs' Calls to Slow Down Draw Media Skepticism(2026-09-14, 4 posts)
- Episode 9: Anthropic CEO Dario Amodei Calls AI Progress a "Warning Sign," Drawing Backlash(2026-09-15, 12 posts)
- Episode 10: Palisade Director Tells BBC the World Is Waking Up to AI Risks(2026-09-15, 2 posts)
- Episode 11: Princeton scholars publish 13,000-word essay arguing AI safety should focus on control over alignment(2026-09-15, 8 posts)
- Episode 12: Debate Flares Over 'AI Is Slowing Down' Narrative(2026-09-15, 2 posts)
- Episode 13: OpenAI confirms weeks of AI safety talks with Anthropic and Google DeepMind(2026-09-15, 6 posts)
- Episode 14: Altman Expresses Strong Confidence in Industry's Ability to Build AI Safely(2026-09-16, 3 posts)
- Episode 15: NY Post Report Targets Anthropic's Handpicked AI Safety Watchers Over EA Ties(2026-09-16, 5 posts)
- Episode 16: AI Giants' Joint Call for Regulation Slammed as Anticompetitive Ploy(2026-09-17, 4 posts)
Primary sources
- [source] New 13,000-word essay: AI safety should bet on control and governance over alignment — sayashk · 2026-09-15
- Narayanan & Kapoor's 13,000-word essay: pacing the AI frontier without consensus — random_walker · 2026-09-16
6 near-duplicate retellings: sayashk · sayashk · random_walker · sayashk · AndyMasley · SashaGusevPosts