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
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
Useful for comparing how concrete threshold and mitigation regimes differ across frontier actors.
Reading
AI Governance: A Research Agenda
Still one of the clearest maps of the field: alignment, concentration, institutional design, misuse, and global governance.
Reading
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.
Reading
Managing AI Risks in an Era of Rapid Progress
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