Field Guide
A lab has built a powerful model. Releasing its weights would let others study and use it independently; keeping the weights inside the lab would limit that access. Who should decide, and what would justify their choice?
Why do the six approaches disagree?
Take the decision from the exercise Who gets access?.
The fictional Larch Research Institute must decide how to provide access to a capable model for the next three months. Its internal tests show useful software and cyber assistance but do not establish autonomous attacks, and no misuse has been observed outside testing.
The access options
- Publish downloadable weights that anyone can run or modify. Released copies cannot be recalled.
- Admit qualified external evaluators to a secure enclave, with general access through a monitored hosted service.
- Offer a monitored, rate-limited hosted service, keeping weights and deeper evaluation inside Larch.
Original admission
An independent panel admits qualified teams and makes final admission decisions. Larch cannot veto a qualified applicant.
Revised admission
The revised arrangement gives Larch final admission power and a veto over applicants who meet the same criteria.
Unchanged conditions
Staffing, budget, screening criteria, wait times, research permissions and publication rights stay the same. Capability and misuse evidence, and the public-weights and hosted-service options, also stay fixed.
Qualified enclave participants may inspect and modify the model, export approved findings and publish criticism under the existing export terms. Weights remain inside the enclave.
Larch does not know how much new benefit or misuse wider access would produce. No arrangement guarantees safety or useful criticism.
These illustrative arguments show how the six approaches could read this decision. Several may support the same arrangement for different reasons. To make your own decision, open the exercise.
Publishing the weights would let anyone run or modify the model, and Larch could not retrieve the released copies. An enclave gives qualified outsiders access without distributing those weights. A precautionary argument favors keeping that control while further testing examines the cyber risks and what would justify wider release.
Strategic Competitor
Who else gains access?
Publishing weights would let competitors build on Larch's work. Hosted access or an enclave would keep the weights inside, while offering different opportunities for outsiders to investigate. A strategic argument asks which researchers should gain access, what they could learn, and how that knowledge might benefit a rival.
Other institutes could use comparable admission criteria and recognize one another's evaluations. The independent panel gives Larch less discretion to exclude qualified applicants. A coordination argument asks whether that arrangement could support cooperation between institutes, and how a developer veto would affect their willingness to rely on it.
Democratic Guardrailist
Who can be held to account?
Evaluators may publish criticism under both arrangements. The difference is who gets inside: an independent panel admits qualified teams in the first; Larch can veto them in the second. Public accountability raises another question: who appoints that decision-maker, oversees its choices and hears a complaint from an excluded researcher?
The enclave requires people to screen applicants and support their work. A state-capacity argument asks whether the responsible bodies have the staff, money and authority to do that job. Before choosing, it would need to know how much work each arrangement requires and who can deliver it.
Open Ecosystem Builder
Who is left outside the qualified group?
Publishing the weights lets researchers outside the qualified group reproduce and modify the model. An independent admission panel still leaves some researchers out. Wider access would reduce that dependence on permission, but Larch would lose the ability to recall the copies. Is access for more researchers worth giving up that control?
Who should be able to inspect, release, restrict, or stop a model?
The same four powers, and where each approach would put them.
| Approach | Inspect and release | Restrict and stop |
|---|---|---|
| Precautionary Steward | Outside evaluators test for dangerous capability before release. Release is staged and gated on those tests. | The bodies that set the thresholds narrow access or pause when warning signs accumulate, even at commercial or geopolitical cost. |
| Strategic Competitor | National security agencies and the state's own technical staff. Release follows what keeps a capability edge under rivalry. | Export controls and limits on diffusion to rivals. A pause by one lab or state does not show that rivals will pause too, so this approach asks what the edge costs while they do not. |
| Coordination Architect | Shared verification and evaluation frameworks that cross borders. Release under rules that travel between firms and states. | Common standards and incident mechanisms agreed across rival blocs. A stop by one state or one lab does not by itself create a shared rule that others recognize. |
| Democratic Guardrailist | Publicly answerable bodies with standing to audit. Release through a process the public can contest. | Accountable institutions, with rights protections on deployment and monitoring. Labs, security agencies, and technical insiders do not get the last word. |
| State Capacity Builder | Staffed regulators with audit, reporting, and procurement powers. Release depends on whether the state can supervise what it approves. | The same bodies, where they exist. A restriction the state cannot enforce has an uncertain effect, and this approach asks who would carry it out before counting on it. |
| Open Ecosystem Builder | Anyone with access. Broad release lets many parties examine and reproduce the work. | Checks on gatekeepers and closed bottlenecks. A stop imposed by incumbents on safety grounds is a claim to be tested. |
Where nearby approaches part ways
The same proposal can appeal to two approaches for different reasons. These comparisons show where those reasons diverge.
Democratic Guardrailist and Precautionary Steward
Shared ground: Both want a stronger check before broad deployment than a lab's own safety case.
They disagree about what the check answers to. The Steward binds it to the capability itself: cross a threshold and the burden of proof rises, whoever is in charge. The Guardrailist binds it to public oversight: a check the public cannot contest is a private arrangement, however cautious.
Strategic Competitor and State Capacity Builder
Shared ground: Both treat AI governance as a problem for states, and both distrust rules that exist only on paper.
The Competitor asks what keeps a capability edge under rivalry and reads institutions as instruments of that contest. The Capacity Builder asks who can supervise, procure, and verify in practice, and what an edge is worth to a state that depends on the firms it is supposed to govern.
Open Ecosystem Builder and Coordination Architect
Shared ground: Both worry that frontier capability is concentrating in a few labs and a few states.
The Ecosystem Builder answers with access: put models and tools in more hands, so no gatekeeper decides what can be studied or built. The Architect answers with rules: cross-border standards and verification that outlast any one lab's advantage. Each sees a cost in the other's cure. Wider access allows more researchers to reproduce the work, while Larch loses the ability to recall released copies. Shared rules raise a different concern: whether the participating firms and states would exclude others.
Coordination Architect and Democratic Guardrailist
Shared ground: Both want enforceable rules and distrust voluntary self-governance by labs.
The Architect looks outward: legitimacy comes from standards and verification that rival states accept. The Guardrailist looks inward: legitimacy comes from domestic institutions the public can hold to account. They may back the same rule for a model. They part over what makes it legitimate, a standard agreed among states or a decision a domestic public can contest.
The axes behind a result
Your answers give positions on the axes below. Each approach weights those positions differently; the comparison determines which result leads. Method
Risk horizon
Do you prioritize current harms first, frontier catastrophic risk first, or insist on holding both in view?
Present harms firstFrontier risks firstDeployment pace
Do you learn through release and iteration, rely on threshold guardrails, or favor stronger precaution before deployment?
Iterate through useSlow at thresholdsPublic oversight
Should labs mostly govern themselves, or should states and public institutions supervise more directly?
Lab-led governancePublic authorityCompetition vs coordination
Do you see rivalry as the binding constraint, or coordination as the only durable answer?
Coordination firstCompetition firstOpenness vs control
Should advanced systems diffuse widely, or should access remain controlled and staged?
Broader accessControlled accessMilitary role
How bounded or how extensive should defense and military uses of frontier AI be?
Civilian restraintBounded defense useLegitimacy and rule-setting
Who should decide: technical insiders, governments, publics, or multilateral institutions?
Technocratic discretionPublic legitimacyHuman future
Do you treat transformative AI futures as promising, dangerous, constrained, or morally ambiguous?
Transformative futuresHuman control
What the questionnaire leaves out
Six ways of arguing about AI that the six approaches touch without scoring. Each ends with a question to put to your own result.
Labor, automation, and bargaining power
AI governance is also workplace governance. Who captures the productivity gains, who is displaced, and whether workers have any say over deployment are governance questions before they are economic ones.
In the questionnaire: The questions touch present harms and human control. They do not ask about unions, labor law, or distributional conflict at work.
How would you judge workplace AI if workers' bargaining power were your main concern?
Surveillance and platform power
AI systems can deepen monitoring, prediction, and behavioral control by firms and states long before any frontier-risk question arises.
In the questionnaire: The questions cover oversight and legitimacy. They do not treat surveillance, policing, or platform governance as traditions with their own answers.
Where does your result treat concentration as a safety problem, and where should it treat concentration as a problem of political power?
Environmental and compute externalities
Compute, data centers, water, energy, minerals, and grid demand are material constraints on what can be built and where.
In the questionnaire: The questions register capacity and control instincts. They do not ask about climate, energy, or the politics of extraction.
Would your threshold for deployment move if the compute cost were visible on every capability decision?
Global South dependence and sovereignty
Most states will meet advanced AI through imported cloud, foreign model access, compliance costs, and infrastructure they do not own. Few will ever choose a frontier lab's release policy.
In the questionnaire: State Capacity Builder and Open Ecosystem Builder point at this problem. The questions do not ask about development strategy or dependency.
Does your result assume the actor has the capacity to govern, or ask what happens when it does not?
Chinese official framings
Chinese governance language combines sovereignty, development, safety, state capacity, and multilateralism in ways that do not map onto the categories of the U.S. debate.
In the questionnaire: The reading lists include Chinese and global-governance sources. None of the six approaches is built from Chinese official doctrine.
Which parts of your result are about governing AI, and which are about the political vocabulary you are used to reading?
Open-source and anti-incumbency politics
Open access can be a technical practice, a market strategy, a civil-liberties claim, or a challenge to concentration in a few labs and states.
In the questionnaire: Open Ecosystem Builder carries this concern as one position. The questions do not separate its security, industrial, and anti-monopoly strands.
When you defend openness, are you defending innovation, accountability, sovereignty, competition, or all of them at once?
Where to read next
Four readings that every approach page starts from. Each approach page adds its own first reading and its strongest critique.
Allan Dafoe · 2018
AI Governance: A Research Agenda
Still one of the clearest maps of the field: alignment, concentration, institutional design, misuse, and global governance.
Independent international expert group · 2025
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 · 2019 / updated 2023
OECD AI Principles
A high-level intergovernmental baseline for trustworthy AI, rights, robustness, and accountability.
NIST · 2024
NIST AI RMF: Generative AI Profile
Shows how governance looks when translated into operational risk-management language.
Precautionary Steward · Strategic Competitor · Coordination Architect · Democratic Guardrailist · State Capacity Builder · Open Ecosystem Builder
AI Futures
The six approaches argue about the next release. AI Futures asks about the world after many releases: nineteen possible futures, from continued human rule to authority passing to artificial systems, and the conditions you would want in each. Browse the whole AI Futures catalogue or say which conditions you would want. AI Futures is a separate exercise and starts empty; it does not change an AI Governance result.