Princeton professors Arvind Narayanan and Sayash Kapoor published the essay “A big-tent or small-tent AI safety movement?” on October 1, 2026, in their Substack “AI as Normal Technology,” which has 87,000 subscribers. In it, they reject both main versions of the debate on AI existential risk: the thesis that catastrophe is real and inevitable, and the mirror thesis that warnings about x-risk are a psyop or regulatory capture. The authors' own position: the warners are sincere but wrong, and the fear framework is counterproductive for AI safety itself. Instead, the essay proposes a measurable agenda: system transparency, legal liability for negative externalities, embedded evaluation, and a possible ban on autonomous recursive self-improvement. The publication comes against the backdrop of the sharpest split in the AI safety community in recent years and formulates an alternative to both camps.

image

What happened

The essay “A big-tent or small-tent AI safety movement?” was published on normaltech.ai — this is the Substack “AI as Normal Technology,” which Narayanan and Kapoor run together. Narayanan is a Princeton professor and the author of the term “AI as normal technology,” and Kapoor is his long-time co-author on model evaluation topics. The authors do not deny the seriousness of the threats as such: they agree that society underinvests in resilience to catastrophic risks, and support this with examples. A 2019 WHO and World Bank report warned of a pathogen capable of killing 50–80 million people, at a preparation cost of $1–2 per person per year. Recommendations from more than eleven high-level commissions, issued since 2009, have not been implemented. Lloyd's of London in 2022 refused to insure severe cyberattacks supported by states. After publication, Narayanan was the first to post the essay on Hacker News under the nickname randomwalker.

Context

To understand the significance of the essay, the background of 2026 is needed: the debate on existential risk was set by Coxon's resignation, disclosed private p(doom) of AI company leaders, and the agent-based hacking attacks of summer 2026, which were written about by NYT and WSJ. The authors' framework, “AI as an amplifier of existing risks,” shifts attention from hypothetical superintelligence to what systems are already multiplying today: cyber risks, bio risks, and so-called cumulative risks — in the terminology of Atoosa Kasirzadeh. This leads to a different view on model evaluation: instead of an alignment framework for AI-enhanced bio risk, it is proposed to measure dual-use capabilities in specific scenarios. Private p(doom) at the same time remain unfalsifiable quantities: they have no methodology, calibration, or reproducibility, making it impossible to rely on them in policy, and the essay does not propose this. Another idea of the text is to involve the skeptical cybercitizenship in the spirit of CISA in safety work, as promoted by commentator Cory Scott, instead of keeping the coordinated safety community in isolation.

Why this matters for the industry

For the industry, the essay is not a product or standard, but a ready-made formulation of an alternative narrative framework, against which companies and their advisors can check positioning and regulatory expectations. The authors highlight two clusters with the greatest reserve of agreement — transparency and liability, and call embedded evaluation the methodological core: model evaluation embedded in real products and workflows, not just in isolated benchmarks. This addresses the long-standing gap between benchmark results and system behavior in deployment. The general signal is that influential academic voices are pushing the regulatory focus from superintelligence to measurable catastrophic and cumulative risks. Nothing changes in production today, but the vocabulary for different work is already available: inventory of what exactly your system amplifies, regression evaluations when releasing agents, and logging actions — all of this is cheap and useful in itself. If the wave around the essay continues, eval gates in release processes, standardization of agent traces and incident processes, “safety dashboards,” and transparency and liability checklists in procurement requirements are likely.

Why this matters for users

For readers, the essay provides a compact map of the split in the AI safety movement, which from the outside seems like a united front. The big-tent and small-tent debate is a debate about strategy: “big tent” means pluralism of values, a positive vision instead of intimidation, and recognition of the gap with public opinion, given that according to Gallup data, only about 0.5% of those surveyed consider AI the main problem. Such a map works as a lens: reading the next statement about AI risks or regulation, one can ask whether a measurable mechanism is being proposed — transparency, accountability, embedded evaluation — or another mobilization of fear. The primary source is fully open on normaltech.ai, and arguments can be discussed and verified in the Hacker News thread started by the author himself.

What is still unknown / limitations

The essay expresses the position of two academic authors, not a testable prediction: confidence that this agenda will almost certainly become regulatory requirements should be considered an overinterpretation. Community response at the time of material preparation was just beginning: the Hacker News thread contained 1 point and 0 comments. The only hard measure — a possible ban on fully autonomous recursive self-improvement — is not operationalized: without a formal definition of “autonomy” and verification methods, such a ban is technically unenforceable. Scenarios about shifting the vocabulary of public discussion, standardizing eval gates and public reporting — these are cautious forecasts, not completed changes; whether the big-tent framework will take root at all will be shown in the coming months.

Sources

Author

Look at AI, editorial team