The Powers Map

What powers must exist to govern catastrophic AI risk, who holds each one today, and what should constrain it?

This map is Stage One of the founding inquiry. Before anyone designs an institution, the inquiry needs an inventory: which powers governing catastrophic AI risk requires, who already holds them, where they are missing, and which must never sit in the same hands. Government without territory advances only if the map shows a gap that existing holders cannot fill.

The map sorts powers into four families. Epistemic powers establish what is true. Decision powers determine what should be done. Consequence powers make a decision matter. Legitimacy powers determine whether the whole system deserves acceptance.

The assessments of current holders and gaps below are SIRDA editorial hypotheses. Each is open to objection, and holders named here can correct the record through their Ecosystem profiles.

Epistemic powers

Capability evaluation

The power to test what a system can do under stated conditions.

Held today by developers' internal evaluation teams, government institutes such as the UK AI Security Institute and the US Center for AI Standards and Innovation, and independent evaluators such as METR and Apollo Research.

Gap. Independent evaluators generally work by the developer's invitation, on the developer's schedule, often with the developer's funding. No evaluator has guaranteed access to frontier systems across borders.

Constraint. Evaluation funding should be pooled and evaluators assigned, so no developer pays the assessor it chooses. The credit rating agencies show what happens otherwise (Chapter 9).

Related questions: What activates jurisdiction?; How should disputed technical evidence become a decision?

Incident detection and reporting

The power to learn that something has gone wrong and to tell the people who need to know.

Held today by developers, national security and infrastructure agencies such as CISA, and public incident collections such as the AI Incident Database.

Gap. No authenticated channel carries incident warnings between developers, evaluators, and governments across borders.

Constraint. A reporting channel sits in some jurisdiction and can be compelled by it, as SWIFT was (Chapter 9). Its governance must say what happens when a host state interferes.

Related questions: What limits would apply during an emergency?; How would this institution relate to existing states?

Compute and hardware measurement

The power to verify how much computation a system used and where.

Held today by cloud providers and chip manufacturers, which see usage directly, and researchers such as Epoch AI, which estimate it from public information.

Gap. No party verifies training compute independently of the developer and its cloud provider.

Constraint. Measurement opens review; it never decides the outcome. Compute is the one auditable primitive, and it is located, so whoever measures it has territorial exposure.

Related questions: What activates jurisdiction?; Who may change the jurisdictional thresholds?

Decision powers

Setting thresholds

The power to decide which systems and activities fall under review.

Held today by legislatures, such as the EU through the AI Act's compute presumption and its delegated revision powers, and by developers through their own safety frameworks.

Gap. Thresholds differ by jurisdiction and by company, and no procedure reconciles them across borders.

Constraint. The body that proposes a threshold change must be separate from the bodies that authorize and review it. Whoever sets thresholds sets the institution's own scope.

Related questions: Who may change the jurisdictional thresholds?

Adjudicating disputed risk

The power to decide, binding on the parties, whether a contested development may proceed.

Held today by no one across borders. Within a country, a regulator or court may decide. Between countries, Plan A escalates unresolved disputes to heads of state.

Gap. This is the missing procedure the inquiry exists to examine.

Constraint. The adjudicator must be separate from the evaluators who supply evidence and from those who carry out consequences. Parties should help select the panel, as in commercial and sports arbitration.

Related questions: Who resolves a dispute when powerful participants disagree?; How should disputed technical evidence become a decision?

Determining equivalence

The power to decide whether another regime's oversight meets a shared standard.

Held today by each regulator for itself. Under the Basel home-host model, a host supervisor judges a foreign bank's home supervision unilaterally.

Gap. No reciprocal procedure exists for AI oversight regimes.

Constraint. No party may certify itself, and no single state or bloc may decide for everyone.

Related questions: Who determines equivalence?

Emergency action

The power to act before full review when delay could be catastrophic.

Held today by developers, which can pause their own systems, and by states under domestic emergency law.

Gap. No body can order interim measures across borders.

Constraint. Every emergency measure needs a defined scope, an expiry, and independent review that can still end it. The sports ad hoc division shows binding decisions within 24 hours are possible (Chapter 9).

Related questions: What limits would apply during an emergency?

Consequence powers

Jurisdictional power

The power to bind actors that have accepted an institution's authority.

Held today by no AI governance body across borders. Developers' own safety frameworks are self-imposed, and binding obligations come from national or EU law.

Gap. The hypothesis proposes to create this power through accession.

Constraint. It reaches members only. Authority over anyone else needs a separate basis.

Related questions: Where would legitimate authority come from?; How do withdrawal and functional defection work?

Access conditioning

The power to condition access to something valuable on meeting a standard.

Held today by cloud and compute providers, chip suppliers, marketplaces and app stores, enterprise and government customers, investors, and insurers. Gene synthesis providers already condition orders on screening, the closest working precedent.

Gap. Holders apply their own criteria, with no shared standard and no route for an outsider to prove equivalent safety. The Ecosystem has no insurers yet, a holder class to map next.

Constraint. The six safeguards on the cartel question: state authorization where possible, public standards, an affordable equivalence route, proportionality, independent review, and access that depends on meeting the standard, with membership as one way to show it.

Related questions: How can decisions reach actors who never joined?; Can infrastructure restrictions avoid cartel power?

Sovereign power

The power of states to license, control exports, set procurement rules, and assign liability.

Held today by national governments, each within its borders. Chip export controls are the most consequential current use.

Gap. States use these powers on their own findings, with no shared evidence or procedure.

Constraint. Where a consequence depends on law, the government constitutionally entitled to make that law must authorize it.

Related questions: How would this institution relate to existing states?; How should military autonomy fit within the mandate?

Legitimacy powers

Representation and authorization

The power to authorize a governing body on behalf of the people it affects.

Held today by legislatures and electorates, each for its own country.

Gap. People exposed to cross-border AI risk have no shared route to representation.

Constraint. Agreement among developers cannot substitute for it (FC1).

Related questions: Where would legitimate authority come from?; What rights would affected parties retain?

Public, workforce, and market pressure

The power of the public, employees, customers, investors, and civil society to change an institution's incentives by withdrawing trust, labor, money, or consent.

Held today by everyone the technology depends on.

Gap. This power works in both directions. In November 2023, most of OpenAI's staff threatened to resign unless the board reversed its removal of the chief executive, and the board reversed course within days. Pressure that can check a governing class can also defeat the oversight meant to check a company.

Constraint. Public legitimacy should shape values, representation, risk tolerance, and whether an institution keeps its mandate. Whether a system crossed a capability threshold is a technical fact, settled by independent evidence.

Related questions: Where would legitimate authority come from?; Can the implementers prevent capture, including their own collusion?

Review and appeal

The power to correct a governing body's mistakes.

Held today by domestic courts, for domestic decisions.

Gap. Nothing reviews cross-border decisions about AI risk. In sport, the Swiss Federal Tribunal reviews Court of Arbitration awards on narrow grounds, and the European Court of Human Rights has held compulsory sports arbitration to full fair-trial standards.

Constraint. The body that acted may never review its own action.

Related questions: What rights would affected parties retain?

Amendment and withdrawal

The power to change the rules, including the rules about who changes the rules, and to leave.

Held today by no one for a body that does not yet exist.

Constraint. An institution that can enlarge its own mandate fails FC5. Exit must end future participation without erasing obligations already incurred.

Related questions: Who may change the jurisdictional thresholds?; How do withdrawal and functional defection work?

Powers that must stay apart

Counter-agency sets the rules for this map. Some powers may never sit with the same body:

Evidence and decision. The evaluator who finds a danger does not decide the response.

Decision and consequence. The body that decides does not carry out the consequences it orders.

Emergency action and its review. Whoever acts in an emergency does not judge whether the action was justified.

Threshold setting and its review. Whoever proposes to move the boundary does not approve the move.

Funding and evaluation. Whoever is assessed does not pay or choose the assessor.

Help complete the map

Each power above should eventually link to every holder in the Ecosystem and to the evidence for each gap. Organizations can correct how they appear, name powers they hold that the map misses, or dispute a gap. Challenge the map.