How Do We Act Together?

A justified response may depend on evidence held by one institution, authority exercised by another, and action carried out by a third. Each can perform its own work competently while the consequence that joins them remains unattended. The expectation developed in the previous chapter—that a response should take effect in time—therefore reaches beyond any single organization. We need the capacity to govern together.

Computational governance gives this expectation a further direction. Institutions could deliberately develop ways to follow consequential relationships across their boundaries, examine possible responses, and check what their separate actions accomplish together. Governing capacity would become a shared subject of research and development, alongside the technology whose consequences it must address. That is the collective promise of the proposed paradigm shift.

Its achievement would require more than connecting information systems. Institutions have different duties, powers, resources, and interests. Better connections can help them act, but can also concentrate influence or carry an error farther. The question is how to develop a common capability while keeping its exercise legitimate and its judgments open to challenge.

From separate duties to collective capability

Consider a further extension of the hypothetical frontier AI developer discussed in earlier chapters. An evaluation prepared by the developer supports an operator’s permission to use a system under particular conditions. An independent evaluator identifies a possible change in those conditions. The developer can investigate its research, the operator can reconsider its use, and the evaluator can examine and report what its remit allows. No failure or harm has been established. The question is what this finding means for continued permission.

The finding might reach all three institutions without resolving that question. The developer could complete its inquiry into the changed conditions. The evaluator could report uncertainty accurately. The operator could confirm that its original approval was properly obtained. Yet nobody’s completed task would necessarily establish whether the evaluation still supports the permitted use. That relationship requires attention of its own.

This is where the manifesto’s first priority becomes demanding:

We value collective capability over procedural compliance.

Procedures are a genuine achievement. They make duties predictable, preserve records, support consistent treatment, and restrain arbitrary decisions. Cooperation would be harder if every institution invented its responsibilities afresh. A dependable procedure can also protect someone who lacks the influence to obtain informal help. We should want these benefits to survive.

The priority asks what the procedures collectively enable. When a permission depends on a finding made elsewhere, completing separate reviews is insufficient unless their relationship can be examined. A gap of this kind should be repaired through legitimate arrangements. It is no invitation for a participant to disregard its obligations or claim another institution’s powers.

Sharing information can make the finding available. An agreement can identify who must assess its implications. Compatible arrangements can let a justified restriction in one institution take effect in another. A shared institution can hold resources or powers that separate participants cannot adequately supply. These arrangements accomplish different work, and any of them may be needed. Their purpose is to create a capacity that can actually address the consequence, rather than to demonstrate that cooperation has occurred.

The shift toward computational governance carries that purpose into the development of governing methods. We can ask whether a changed condition becomes visible to those responsible for a dependent permission, whether they can understand its significance, and whether their responses cover the activity in question. Those connections can themselves be investigated and tested. Meetings and agreements remain necessary; they gain a more exacting object of work.

Following consequences across boundaries

In the hypothetical example, the relevant connection is between an evaluation’s conditions and an operator’s permission. Suitable computational methods could help maintain that connection as the evidence changes. A finding could be brought to the attention of those responsible for the permission, together with the grounds needed to assess it. Where some records must remain protected, the arrangement could support a justified examination without giving every participant unrestricted access.

This would require institutions to represent enough of their governing work for the connection to be useful. A bare approval may reveal little about the conditions on which it depended. A finding detached from its uncertainty may convey too much confidence. Making approvals and findings available with their conditions and uncertainties in a usable form is substantive work: deciding what matters, producing evidence, checking it, and keeping the account current. Computation cannot recover a condition that nobody examined or recorded.

The opportunity also extends to anticipating consequences. Participants could examine how a proposed restriction on the developer’s activity interacts with a change in the operator’s use. They could compare possible responses under stated assumptions, including what remains exposed when only one institution acts. AI might assist that inquiry; conventional software or simpler methods might be sufficient. The useful choice depends on the task and the evidence of performance.

Once the relevant actors have authorized a response, computational assistance could help carry it out across their separate responsibilities and establish what took effect. A restriction might be recorded at one point while a dependent activity continues elsewhere. Checking the combined result matters because the intended protection may depend on the connection. An acknowledgment that instructions were received would not establish that protection.

Three questions help keep the ambition grounded:

These questions concern connected governing work. A technically correct finding could reach someone who has no power to act. An authorized response could arrive after its useful window. A joint analysis could exclude the people whose experience would reveal the consequence that matters. Each possibility directs attention to the arrangement as a whole, including its authority, resources, and sources of knowledge.

Institutions need not agree on everything to work in this way. They may share an observation while disagreeing about its significance or the exposure they consider acceptable. They may have different lawful responsibilities. A common way of examining evidence should make those differences clearer, rather than erase them to produce a single answer. A shared computational facility can support several judgments; reliance on it does not make it the rightful judge.

Connected governing capacity deserves deliberate investigation and investment. The standard is what the participating institutions can adequately govern together, with their separate duties and powers still visible.

Governance Labs

Where could institutions develop that capacity together? Some of the work requires access to different kinds of expertise and evidence. Some requires examining interactions that no participant can reproduce alone. Some requires an independent participant to challenge an apparently successful result. We need settings in which these contributions can meet before consequential reliance makes a weakness costly to discover.

This book proposes Governance Labs: institutions dedicated to developing and testing the capacity to govern consequential activity. They would bring developers and operators into inquiry with public overseers, independent researchers, and representatives of people affected.

The aim would be to investigate governing arrangements, including suitable computational methods, and establish what those arrangements can and cannot adequately accomplish. This is a proposal for their purpose, not a claim that a particular institution has already achieved it.

Governance Labs would advance what we can call the Governance Frontier: the developing boundary of what institutions can adequately examine, authorize, constrain, and correct together. The frontier concerns governing capability. The frontier AI developer in our hypothetical builds AI systems; a Governance Lab would investigate the means of governing such activity. The two institutions have different jobs, even when the developer contributes to the inquiry.

The Governance Frontier

More capable technology creates possibilities for action. More capable governance creates possibilities for justified action. Developing the second can open worthwhile uses that would otherwise remain unjustified, as well as reveal where limits are necessary. Advancing the Governance Frontier means demonstrating that institutions can govern a consequential activity at an appropriate scope. A sophisticated tool or a busy laboratory does not establish that achievement.

Return to the independent finding in the hypothetical. A Governance Lab could use synthetic records or a simulation to test whether the finding reaches the people responsible for the affected permissions. Participants could examine whether the grounds are usable, whether uncertainty survives the transfer, and whether the recipients can respond within the relevant time. They could then test whether separately authorized restrictions address the interaction and whether their effects can be checked.

An exercise could reveal a gap even when every computational component performs its assigned task. Suppose the finding reaches the correct recipient promptly, but that institution cannot require the necessary change. Faster transmission would leave the missing power unresolved. The lab could identify the decision that participants or public authorities need to confront. It could not supply that power by completing another test.

Other results might show that a proposed method makes little useful difference, or that a simpler arrangement performs better. Those are valuable findings if they are allowed to count against the proposal.

The comparison should include the actual existing method and feasible alternatives, together with the work required to prepare evidence, verify results, and maintain the connection. Reducing one institution’s reporting effort while increasing everyone else’s review burden would need an explanation.

Testing should begin in settings justified for that purpose. Simulation and synthetic records can permit useful inquiry without exposing people to a consequential experiment. Some questions may require bounded operational tests, but a research ambition alone cannot justify their burdens. The relevant authority must consider the proposed exposure and protections.

Results from a controlled setting also leave a further question: whether the arrangement remains adequate in the operation, scale, and conditions for which it is proposed.

Adoption needs investigation too. A method may work in the exercise yet have no authorized owner, funded staff, or accepted place in ordinary review. A lab could examine what prevents its use and which changes would make it practicable. Demonstrating that barrier gives the participants a question to resolve; it does not grant the lab powers to resolve it for them.

The labs would need to examine the production of evidence as well as the tools that process it. Independent findings, missing information, disputed interpretations, and unsuccessful interventions should have a place in the inquiry. Public authorities and affected participants must help shape what is being tested and what would count as success. Otherwise the research could become an elaborate demonstration of whatever the most influential participant already wishes to show.

The institutional form may vary. A lab could be established through an existing public body, an independent research institution, or a legitimately constituted shared arrangement. Its mandate would determine what it may examine and do. Research findings could inform revised permissions or governing methods through the authority responsible for them. The laboratory name confers no power to authorize deployment, rewrite obligations, or certify all uses as acceptable.

Sharing capability and concentrating power

The attraction of a common facility is real. Smaller institutions could gain access to expertise, methods, and evidence they cannot sustain alone. Joint work could make an expensive examination feasible and help participants understand consequences beyond their own field of view. People who currently face an institution’s fragmented responsibilities could benefit from a response that connects those responsibilities properly.

These possibilities give operational convenience a serious claim. Repeated work consumes resources that could support useful scrutiny. Dependable common methods can make participation easier and free attention for difficult judgments. Where those benefits are demonstrated, integration may expand meaningful choice rather than diminish it.

We value meaningful choice over operational convenience.

The tension arises when the convenient arrangement becomes difficult to question or replace. A common tool can embed choices about what evidence matters, whose concerns receive attention, and which consequences appear in the comparison. If participants cannot inspect those choices or pursue another inquiry, the facility may shape their judgments more deeply than its formal mandate suggests. Its influence deserves attention even when final decisions carry separate signatures.

Greater access is not the same as greater influence. A less-resourced participant may be allowed to attend while lacking the time or expertise to challenge the research agenda. Participation costs include preparing evidence, understanding methods, and responding to findings. Funding and assistance should make those contributions practicable. The aim is to strengthen the participant’s ability to exercise its responsibilities, including when its judgment differs from that of the group.

Who pays for a Governance Lab also matters. Developer funding might make valuable inquiry possible while giving the developer influence over which questions are asked or which findings become public. The arrangement must be able to sustain inquiry and appropriate scrutiny of findings when the funder objects or threatens to withdraw support. The arrangement must enable inquiry that can reach unwelcome conclusions, preserve the grounds for examining them, and support scrutiny by people who do not depend on the developer’s approval.

Shared analysis can also become a common blind spot. Participants might accept the same mistaken account because it appears consistent across their institutions. Adding another AI system would not by itself establish an independent perspective.

Independence requires room to investigate differently, access appropriate to that inquiry, and a route by which contrary evidence can matter. Workers, evaluators, and affected people need protected ways to raise concerns outside the shared account.

Protection of information remains part of the design. Cooperation does not entitle every participant to every record. Confidential research, personal information, and protected reporting require justified access. A joint method must show that its conclusions can be examined by the appropriate people while those protections survive. If that cannot be achieved, the resulting limit on knowledge must remain visible in the decision about permitted activity.

These requirements consume resources and can slow some work. That cost cannot always be removed by better software. Independent examination may duplicate effort; a protected route may complicate integration; preserving a different method may reduce convenience.

The priority for meaningful choice asks when those costs preserve a capability or right worth retaining. It does not excuse ignoring common duties or leaving others to bear an avoidable consequence.

Acting together amid legitimate differences

There is a strong objection to a distributed arrangement: too much independence can prevent effective action. Participants may wait for one another, select only the convenient parts of cooperation, or decline burdens while benefiting from others’ work. Some consequences cannot be governed adequately through voluntary exchanges alone. A common capability may need binding commitments and authority to require action.

The answer must take that objection seriously. Participants need to establish what each must contribute, how responsibilities connect, and which decisions require legitimate common authority. A shared institution may be warranted where separate powers cannot address a joint consequence. Conflicting mandates need a legitimate route to resolution where one exists; more evidence cannot by itself settle them. Where common authority is absent, a test can clarify the gap and possible responses, but the political decision remains. It must be made through arrangements entitled to make it.

Cross-border cooperation brings this distinction into sharp focus. Institutions in different jurisdictions can examine evidence together without having interchangeable powers. A finding that justifies action by one authority may require a different assessment elsewhere. Differences in rights, acceptable exposure, and public priorities may persist after uncertainty is reduced. Computational analysis can clarify the consequences of those differences; it cannot decide whose interests should prevail.

Inherited boundaries can also shape which consequences receive attention. If reporting duties and review resources are organized around activity inside each jurisdiction, an effect spanning them may have no equipped reviewer. Cooperation should identify that gap without treating every difference as a defect. Some differences protect legitimate choices about rights and acceptable exposure.

Evidence can sometimes travel usefully across those boundaries. It needs its system version, operating context, test conditions, uncertainty, and provenance—the account of where it came from and how it was produced. A recipient must judge whether it answers the local question; retesting may be necessary. Reusing evidence does not require accepting another institution’s conclusion, treating standards as equivalent, or transferring legal powers. Compatibility can reduce avoidable duplication while leaving decisions with the authorities entitled to make them.

People outside the participating institutions may still bear burdens. They need appropriate representation and routes to challenge. Collective capability must serve legitimate purposes; coordination cannot justify an unfair distribution of costs.

Participants may preserve different judgments while committing to a common restriction or agreeing on evidence needed for reconsideration. Such commitments need justification within their powers. Where cooperation cannot make an activity adequately governable, participants must establish who has the powers, resources, and means to strengthen capacity or narrow permission. Missing powers require legitimate institutional reconsideration; hoped-for capacity cannot support present permission.

Governance Labs would give deliberate collective development a place to happen. Their contribution would be demonstrated capacity and a clearer account of its limits. The larger expectation reaches beyond any institutional name: those who advance consequential technology should help develop the means by which it can be responsibly governed, and those entrusted with oversight should be able to investigate those means independently.

Acting together can expand capability and create dependence on shared evidence, methods, and facilities. The next question is which practical choices must remain available as our ability to govern depends on arrangements we share.