AI Governance

Precautionary Steward

Treat frontier AI as a system that may warrant slowdown, stronger thresholds, and demonstrable safety before broad deployment.

Definition

The argument

The burden of proof should rise with capability, especially where loss of control, catastrophic misuse, or systemic dependence could follow.

Institutional priorities

  • credible thresholds and eval gates
  • stronger deployment restraint under uncertainty
  • institutional willingness to pause when warning signs accumulate

Risks it foregrounds

  • premature deployment becoming irreversible
  • capabilities outrunning governance capacity
  • elite incentives overwhelming safety warnings

Strongest critique

The strongest objection

Your critics will say that precaution can become an incumbent-friendly politics of delay, especially if the institutions doing the slowing are not broadly trusted.

Critique to start with

On the Dangers of Stochastic Parrots

Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, Shmargaret Shmitchell · 2021

The canonical present-harms critique of scale-first language-model development.

Question to sit with

The live tension

How do you slow or narrow diffusion without simply hardening concentration and calling it safety?

International lens

More open to international coordination and shared thresholds, but often skeptical that rivalry will naturally produce safe outcomes.

Core disagreement

The disagreement

The closest neighbors share most of the vocabulary. The split usually comes down to a small number of axes where this archetype makes a different call.

This archetype often sits near other cautious, rule-building positions. The main split is whether the center of gravity should be severe frontier risk, public accountability, or transnational coordination.

Nearby

Democratic Guardrailist

Treat AI governance as a legitimacy problem as much as a capability problem, with democratic oversight, rights, and accountability as the anchor.

Open Democratic Guardrailist

Nearby

Coordination Architect

Treat the hardest AI problems as transnational and institution-building problems rather than purely national or purely technical ones.

Open Coordination Architect

Policy applications

What follows in practice

These proposals belong to this editorial category. A similar result does not establish that you support each one.

  • Back stronger external evaluations before broad release when dangerous capability signals appear.
  • Prefer staged deployment, threshold triggers, and visible incident escalation paths over default rapid rollout.
  • Support governance that can slow or narrow access when evidence becomes worrying, even if that imposes real commercial or geopolitical costs.

Questions to examine

Where the argument is tested

These editorial questions explore the assumptions behind the proposals. They are not a report of current events.

Capability evals before deployment gates

Whether external evaluations, red-teaming, and dangerous-capability tests should harden into mandatory release gates rather than voluntary norms.

When a pause becomes centralization

How to slow or stage frontier rollouts without locking governance authority into the same few labs and agencies the slowdown is meant to constrain.

Reversibility of frontier diffusion

How much weight to give arguments that open weights, agentic systems, or biosecurity-relevant capabilities are effectively irreversible once released.

Starting readings

Where to begin

Start with the central argument, then read its strongest challenge. The full shelves are available here without completing an assessment.

Reading

Frontier AI Safety Policies

METR · 2026

Useful for comparing how concrete threshold and mitigation regimes differ across frontier actors.

Reading

AI Governance: A Research Agenda

Allan Dafoe · 2018

Still one of the clearest maps of the field: alignment, concentration, institutional design, misuse, and global governance.

Reading

International AI Safety Report 2025

Independent international expert group · 2025

Useful as a shared scientific baseline for advanced-AI safety debates across countries rather than a single camp's framing.

Reading

Managing AI Risks in an Era of Rapid Progress

Yoshua Bengio et al. · 2023

A broad coalition statement capturing the frontier-risk mainstream case for stronger governance.

Full reading shelves

Read the result from another angle

Use each note to identify the question a source helps investigate. The shelves follow an editorial model category; authors do not endorse a reader result or an illustrative scenario.

Challenge your view

Which assumptions in this reading deserve the hardest challenge? These sources supply competing arguments, not predictions of your views.

On the Dangers of Stochastic Parrots

The canonical present-harms critique of scale-first language-model development.

Harms of AI

A strong challenge to governance approaches that assume more AI or more state support is obviously beneficial.

Atlas of AI

A material critique of AI as an extractive infrastructure, not just a software or safety problem.

Open-Sourcing Highly Capable Foundation Models

Useful when testing openness-first instincts against the strongest misuse and proliferation critique.

Read the profile’s foundations

Shared field maps and baseline texts before you settle into one governance camp.

AI Governance: A Research Agenda

Still one of the clearest maps of the field: alignment, concentration, institutional design, misuse, and global governance.

International AI Safety Report 2025

Useful as a shared scientific baseline for advanced-AI safety debates across countries rather than a single camp's framing.

OECD AI Principles

A high-level intergovernmental baseline for trustworthy AI, rights, robustness, and accountability.

NIST AI RMF: Generative AI Profile

Shows how governance looks when translated into operational risk-management language.

Go deeper

Operational frameworks, official documents, and live policy tools that show how these debates cash out in practice.

Frontier AI Safety Policies

Useful for comparing how concrete threshold and mitigation regimes differ across frontier actors.

Managing AI Risks in an Era of Rapid Progress

A broad coalition statement capturing the frontier-risk mainstream case for stronger governance.

Common Elements of Frontier AI Safety Policies

Synthesizes what labs are actually doing around thresholds, model evaluations, weight security, and deployment mitigations.

Responsible Scaling Policy

A living example of how a frontier lab publicly structures capability thresholds and safeguards.

Updated Preparedness Framework

Useful for comparing another frontier-lab approach to severe-risk measurement and mitigation.

Safety Cases at AISI

Pushes the debate from generic principles toward structured arguments that a system is safe in a given context.

Widen the frame

How these debates look beyond U.S. and European defaults, including Chinese and broader global-governance lenses.

State of AI Safety in China

The best single English-language guide to the diversity and maturation of Chinese AI safety discourse.

Global AI Governance Initiative

Essential for understanding China's official framing around sovereignty, development, safety, and multilateral governance.

Global AI Governance Action Plan

A current official text on cooperation, development, and global-governance architecture beyond a U.S.-centric frame.

Recommendation on the Ethics of Artificial Intelligence

Useful for human-rights, dignity, and global ethics framing that extends beyond frontier-lab discourse.

Routes

Explore the consequences