What Must Governance Be Able to Do?

What governance has to keep up with

Recognizing a concern gives governance something to address. It does not establish that the institutions responsible can address it. They may understand the problem and lack authority to intervene. They may possess that authority while lacking the evidence, resources, or time needed to use it well. Asking what governance must be able to do takes us from a statement of concern to the conditions for a legitimate, effective response.

AI advancement can change those conditions. A more capable system may perform tasks that earlier evaluations did not examine. New permissions may let it act where it previously advised. Connections between systems may give an output consequences beyond the activity originally reviewed. Oversight must remain adequate for what the whole arrangement can do, including what has changed since permission was granted.

The familiar question is whether governance can keep pace. A review that arrives after an irreversible consequence can be too late even if its conclusion is correct. But faster review alone cannot resolve every mismatch. An institution can process submissions promptly while failing to examine the activity that matters. It can shorten the interval between decisions without improving their grounds.

Consider a different extension of the frontier-development illustration in Chapter 1. Imagine a laboratory running concurrent experiments, using AI-assisted evaluations, connecting research systems, and deploying results in several settings. Each team keeps intelligible records. Reviewers receive them on time. Yet a change approved in one experiment affects the conditions under which another evaluation is reliable, and the connection is not adequately examined. This is a hypothetical arrangement, not a finding about a particular laboratory.

The unit of review can itself preserve the gap. Suppose the laboratory’s forms, staff assignments, and approval records were built around separate experiments. Each team becomes proficient at its own submission, while examining a relationship requires work outside the recognized duty. Faster processing can make that established approach more attractive without making the connection examinable. Repair requires attention to the object of review and the authority responsible for it.

This is a question of complexity capacity: the ability to examine consequential activities and their interactions at the scale where they matter. Governance must be able to retain relevant context, recognize connections, evaluate their significance, and act on the findings. The amount of information received is only part of that capacity. What institutions can do with it matters as much.

We will first examine these demands, then the capacities that could meet them. The final question is whose ability to govern becomes stronger, including the ability to challenge those exercising power.

Seeing the parts and understanding their connections

Visibility begins with establishing what consequential activity is occurring. What exists, what changed, and what relevant capability or interaction has appeared? In the hypothetical laboratory, reviewers need evidence of the connection between experiments, not merely confirmation that reports were submitted. Without a way to discover that connection, they may continue to rely on an evaluation whose basis has changed.

Visibility should be appropriate to the responsibility. An independent evaluator and a member of the public do not necessarily need the same records. Private information and sensitive security details require protection. The demand is for evidence sufficient for meaningful scrutiny, with ways to question the operator's account. Universal disclosure would not automatically provide that understanding.

Comprehension asks what the available evidence means. A reviewer might establish that two systems exchanged information while remaining uncertain whether the exchange invalidated a safeguard. The record supports one conclusion; its significance requires further examination. Responsible governance must distinguish an unanswered question from a favorable answer. Where the uncertainty affects permission, it belongs in the decision about permitted activity.

Complexity capacity adds a further demand: review must examine consequential connections as well as individual activities.

A review may need to connect changes in permissions, shared resources, evaluation methods, and deployment conditions across different responsibilities. No single individual needs to know every internal detail. The arrangement must nevertheless make the consequential relationships examinable and ensure they do not disappear between roles.

The problem can persist even when everyone performs a narrow duty competently. One reviewer checks an evaluation; another checks the deployment it supports. If neither has the information or remit to examine how a later change affects both, individual diligence leaves a collective gap. Improving their work may require shared evidence, additional expertise, a different allocation of responsibility, or limits on what can be connected before review.

Adaptation concerns whether the governing arrangement can respond when its basis changes. A dependable method should not be discarded merely because a newer one exists. It should be reconsidered when it no longer answers the consequential question. That may involve changing an evaluation, seeking different evidence, or narrowing a permission while inquiry continues. Improving the questions matters alongside improving the answers.

Authority concerns which decisions an institution is entitled to make and which powers it may legitimately exercise. Exercising those powers requires resources and an effective route to implementation. It may need cooperation from others. An alert delivered to someone who cannot act leaves an important part of governance unfinished.

Pace matters across all these demands. Evidence can arrive too late; analysis can describe conditions that no longer hold; an intervention can follow the effect it was meant to prevent. The relevant timescale depends on what could change and how much time remains for a response. Updating a rule whenever a model is updated would not by itself establish adequacy. Some protections should endure while the methods used to honor them improve.

Timing should therefore differ for reasons tied to the activity. In a separate hypothetical comparison, a developer changes an experimental system’s tool access while a bank uses a stable model for bounded advice. The first change may require examination before its next consequential action; the second may remain adequately covered by scheduled review and attention to material changes. If the bank grants permission for direct transfers, its response window changes too. Conversely, the developer can have enduring obligations that need no daily revision. Actor labels alone cannot determine cadence. The permissions, exposure, reversibility, evidence, and time available for effective response must justify the difference.

Timing also leaves the quality of judgment open. A forecast can clarify consequences without deciding which burdens are acceptable. Different review intervals need reasons about protected interests and actual exposure; differentiation must not become an unexplained exemption.

These distinctions help identify what is missing. They do not certify an institution or imply that every frontier activity has already exceeded adequate oversight. They ask whether the actual arrangement can meet the demands of the activity it permits. Where it cannot, governance must strengthen its capacity, constrain the activity, or do both.

The capacities that make a response possible

What capacities should we strengthen to meet those demands? This book proposes seven, brought together in the Governance Amplifier Framework. An amplifier names an ability governance should develop. It invites us to examine what an arrangement makes possible and what its exercise could diminish.

The questions can be held together without treating them as a sequence:

Understanding — Can the relevant people understand what is happening and where uncertainty remains?

Judgment — Can institutions distinguish desirable, tolerable, and unacceptable outcomes?

Legitimacy and agency — Who authorizes decisions, who participates, and who can challenge them?

Coordination — Can actors act together without erasing legitimate differences?

Effective action — Can an authorized decision actually change what happens?

Learning and adaptation — Can governance detect and correct its own deficiencies?

Stewardship — Does governance preserve long-term interests and future choices?

The demands we have examined explain what governance must cope with. These capacities express what it should strengthen while protecting people's agency and keeping power legitimate. Their value comes from working together. Better understanding can improve a decision, but it cannot supply a mandate. Effective intervention can realize a legitimate choice, but greater power to intervene can also deepen an injustice.

Understanding includes making consequential matters intelligible to those with different responsibilities. Expertise should help others ask a useful question, examine the grounds for a decision, or recognize uncertainty. A clear explanation earns confidence when its claims can be checked. It should also make clear what has not been established. Collecting more information can become intrusive or overwhelm scrutiny; adequacy requires attention to what the information enables.

Judgment concerns what should happen in light of that understanding. Institutions need to distinguish a worthwhile opportunity from an unacceptable burden, and a tolerable uncertainty from one requiring additional limits. Facts may resolve part of the disagreement. Different interests can remain even when the facts are clearer. A forecast of aggregate benefit would not settle whether a particular intrusion or exclusion is justified.

Legitimacy and agency concern the grounds of authority and people's practical influence over its exercise. Affected people may contribute knowledge unavailable inside an institution. They also have interests that deserve consideration even when their contribution adds no technical insight. Participation must be usable, and consequential decisions need intelligible reasons and routes to challenge. A process that rewards only those with extensive time, expertise, or money can weaken the agency it claims to preserve.

Coordination makes divided capacities useful together. In the laboratory illustration, the relevant knowledge and power may sit with different teams and independent reviewers. Beyond a firm, responsibilities may be divided among operators, public authorities, and institutions in different jurisdictions. They need ways to connect matters that cross those boundaries. Cooperation need not produce identical rules or agreement about every value. It must preserve legitimate differences and avoid quietly assigning authority to whoever operates the shared technical facility.

Effective action connects an authorized decision to its intended effect. It includes investigation, repair, restriction, and the restoration of access, as well as permission for useful work. Return briefly to Chapter 2's hypothetical public service. Suppose a reviewer establishes that an AI-directed change obstructed access but lacks authority to restore it. A more accurate report would help explain the problem; it would not restore access. The reviewer needs a working route to someone able to act, who must consider the finding and affected interests and carry out a justified response in time.

A duty needs the means to fulfill it

Assigning responsibility can leave a gap untouched. If the person responsible lacks authority or resources, improved analysis may only describe what they cannot change. A meaningful duty needs practical means and a route for matters beyond the person's mandate.

Those means must themselves be justified. Expanding enforcement power can impose burdens or expose people to arbitrary decisions. Acting effectively includes acting within legitimate limits and remaining open to challenge. The remedy for a powerless reviewer is not automatically unlimited power for the reviewing institution.

Learning and adaptation allow the governing arrangement to detect and correct its own deficiencies. A missed interaction may reveal a weakness in a particular decision, in the evidence requested, or in the way responsibilities were divided. Learning should reach these sources of inadequacy. It can begin with a test, a complaint, or a useful emerging capability before harm occurs. A revision needs a credible basis; keeping a sound protection can be an informed result.

Stewardship keeps attention on lasting interests and future choices. An arrangement might solve a present analytical problem by becoming dependent on a tool that independent reviewers cannot examine or replace. That dependence belongs in the judgment about improvement. So do the benefits of useful help now. Preserving future choices can require investment and slower integration, but speculative future interests must not casually displace urgent present needs.

None of these capacities should be maximized in isolation. Seeking greater understanding through more data can impose excessive surveillance. More coordination can mean excluded outsiders. More adaptation can mean instability. We need reasons for the capacity being strengthened, the limits placed on it, and the consequences for those affected. There is no aggregate score through which a missing essential protection becomes acceptable because other capacities have improved.

Making complexity governable

The complexity-capacity concern gives computational governance a substantial purpose. If consequential activity exceeds what people can adequately examine unaided, insisting that every governing task remain unaided may preserve an ineffective method. Suitable computation can help institutions retain context across activities, compare evidence, examine interactions, and explore consequences. Some assistance may use conventional software; some may use AI.

Developing this capacity deserves deliberate effort alongside developing AI itself. Institutions can investigate governing methods, test their limits, and compare what they make possible. Where evidence and powers are divided, that work must examine the connections between institutions as well as the work inside each one. Chapter 6 develops the computational possibilities; Chapter 8 asks how actors can build and exercise such capacity together.

Consider what would be useful in the hypothetical laboratory. Reviewers may need a way to discover when a change in one activity affects an authorization elsewhere. Computational assistance could help connect the relevant records and direct attention to the relationship. Its value would depend on whether it identifies consequential connections reliably enough for the task, makes its grounds examinable, and reaches someone able to respond.

Memory matters because a current decision can depend on reasons established earlier and conditions that must continue to hold. Analysis matters because the relevant consequence may lie between activities rather than within one report. Simulation can help examine possible interactions before authorizing them. Each can support a governing judgment without supplying proof that every important consequence has been anticipated.

The assistance should improve the whole arrangement's ability to govern. If it reproduces the same omissions as the reports it analyzes, an additional layer may increase confidence without improving scrutiny. If its operation cannot be independently examined, institutions may acquire new analytical power while weakening their ability to question its findings. Claims of improvement need an identifiable basis, including what became better, for whom, and with what remaining uncertainty.

Computation can also assist the examination of governing methods. It might help identify a recurring omission or compare proposed changes. Authority to revise the method still belongs within the governing arrangement. Proposing a change and authorizing it are different acts. An institution should be able to improve how it discovers deficiencies without delegating its right to decide what counts as acceptable governance to the tool making the proposal.

The standard should be adequacy for the activity permitted. Governance need not reproduce the internal complexity of the AI or become an equally capable system in every respect. Preventive bounds can keep some consequential actions within reach of oversight. Reducing permissions or limiting interactions can make an activity easier to examine. Strengthening capacity and improving governability are complementary choices.

Those choices have costs. Restricting an interaction may prevent useful work. Building analytical capability may require resources that could serve other needs. Introducing assistance may add verification and maintenance burdens. Some activities may need to remain narrower until evidence supports expansion; others may warrant substantial investment in governing capacity. A preference for useful development does not settle the choice in advance.

Better analysis should strengthen political judgment by clarifying what a choice involves. It cannot determine whose interests deserve priority simply by forecasting or optimizing more accurately. Deciding how burdens are distributed, what rights must remain protected, and who may authorize a consequential activity requires legitimate public judgment. Treating those questions as technical outputs would conceal a transfer of authority.

Computational governance therefore asks more than whether an institution has adopted advanced tools. It asks whether the tools, people, responsibilities, and constraints together can support legitimate decisions and effective response. The same question must be asked of the AI used to assist governance.

Whose capability becomes stronger

The remaining question concerns the distribution of capability. A regulator with better analysis and stronger intervention powers may still be difficult to challenge. A developer with improved internal oversight may still control which evidence independent reviewers receive. An arrangement can become more capable for those already holding power while leaving others less able to influence what happens to them.

We value collective capability over procedural compliance.

Here the priority asks whose practical ability the procedures support. A dependable record can give an independent investigator grounds to question an account. A review requirement can connect a finding to someone able to act. A usable challenge can help an affected person obtain correction. Their value is realized through those capacities, with lawful obligations and protections retained.

Suppose an institution improves its own analytical work while leaving its challenge process inaccessible. People may receive more carefully reasoned decisions but remain unable to correct a mistaken premise. That would be a genuine improvement with an important deficiency still present. Calling the whole arrangement stronger should not conceal who remains without practical recourse.

Independent inquiry and affected people's experience are part of governing capacity, alongside official expertise. They can reveal what established methods omit and challenge how improvement is defined. They need appropriate access and ways for findings to matter. Sharing capability may involve investment in institutions that have fewer resources, rather than requiring every participant to acquire the same tools or obtain identical information.

This also gives improvement a humane purpose. People should be able to obtain competent help without becoming specialists in the systems that affect them. Institutions can recognize the limits of their current knowledge and still provide a responsible route to inquiry, constraint, or remedy. A narrower permission or a revised method can preserve useful activity while bringing its consequences within reach of scrutiny. Such responses can be worthwhile even when they are costly or incomplete.

We should expect governance to explain why its capacity is adequate for what it permits, what would cause that judgment to change, and who can question the account. That expectation concerns the power to act and the ability to correct its exercise. The next question is how people should participate in those choices, and who has legitimate authority to decide.