How Can AI Help Us Govern?

Help with the work of governing

Computational governance opens a possibility larger than saving effort on individual tasks: we can deliberately develop the capacity to govern as we develop the technologies being governed. Suitable computation can connect evidence to the conditions of permission, examine consequential interactions, support authorized responses, and help establish whether those responses worked. This chapter argues for a shift in how governing capacity is developed and exercised, with its usefulness demonstrated in the settings where institutions rely on it.

Chapter 5 asked what would make oversight credible. Meeting that expectation can demand more than people can reliably examine unaided. A reviewer may need to connect a proposed change with conditions established across many earlier decisions. Receiving each record promptly would not establish that the relationship had been understood. The institution needs a way to make the important connection examinable.

This is the paradigm shift proposed here. Governing questions, the evidence supporting a permission, and the means of carrying out a decision could become connected, testable parts of an ongoing arrangement. A change in activity could then prompt examination of the conditions it affects, while those responsible retain the context needed to judge it. The ambition is governing capacity that can be examined and improved, rather than an assumption that installing AI improves governance.

Automating inherited tasks can leave that ambition unfulfilled. A sophisticated assistant might classify every submission using categories that omit a new interaction, or route every concern through a workflow unable to reconsider permission. Its success at the assigned task could strengthen reliance on the inadequate arrangement. Institutions must be able to examine the categories and connections themselves, including whether another question or method is needed. Retaining human responsibility does not require unaided execution, but changing execution alone does not repair the governing relationship.

Accountability to people does not require people to perform every operation themselves. It requires arrangements through which consequential power remains answerable. Assistance can make a duty more practicable, provided those responsible can examine its contribution, recognize its limits, and respond when reliance becomes unjustified.

This leaves a real choice about the work. A dependable check might be performed with conventional software. Comparing varied accounts or investigating an unfamiliar relationship might benefit from AI. Some questions require expertise or participation that neither can supply. Choosing assistance should follow what the task requires, including what would happen if the assistance were wrong.

We will consider what this shift makes possible, the work computation can support, and the permissions within which it operates. We will then ask how institutions can develop and test that capacity, and who remains responsible when it becomes part of governing power.

Keeping context and examining connections

A governing decision can depend on conditions that must continue to hold. Keeping those conditions available matters as much as preserving the fact that approval occurred. Suitable computation could connect a proposed change to the evidence and reasons supporting an earlier permission, helping a reviewer recognize what needs renewed attention.

Some obligations can also be represented in forms software can check. An arrangement might require a particular authorization before an action proceeds. A check could establish whether that authorization is present and applies to the action. Deciding whether the permission was justified, or whether a disputed obligation has been interpreted properly, may require further judgment. The useful boundary follows the obligation and its consequences.

There is a bounded precedent for structured information in security assessment. NIST's Open Security Controls Assessment Language, known as OSCAL, provides standardized, machine-readable representations of security controls and their implementation for analysis and reporting.¹ This illustrates how information relevant to oversight can be organized for computational use. It does not establish that public judgment can be automated or that the total work of governing has become inexpensive.

AI may contribute where the work involves interpreting and comparing material that does not fit a fixed check. It could help locate conflicting accounts, identify a relevant passage, or propose a connection that deserves examination. The contribution should preserve its supporting grounds and distinguish what the evidence establishes from what the tool infers. A relationship omitted from the available material may remain invisible to the assistance too.

Return to Chapter 3's hypothetical laboratory, where concurrent experiments have separate records and a change in one activity may affect an evaluation supporting another authorization. Imagine an assisted reviewer identifying that possible connection and pointing to the relevant records and earlier condition. The reviewer can now examine a question that might otherwise have been difficult to find among the separate accounts.

The next work is to check the connection. Did the activities share the condition the tool identified? Does the change affect the evaluation's reach? Is further evidence needed?

The assistance may have found something important or misunderstood the records. Its suggestion does not establish that a safeguard failed. It makes the grounds of a consequential inquiry available for examination. This remains a hypothetical possibility, not a result achieved by a particular system.

Preparing that evidence also involves work inside the governed organization. An operator might use assistance to organize records, connect a claim to its supporting material, or explain the same finding appropriately to different recipients. An overseer might use assistance to compare the submission with other evidence and pursue a missing question. Both contributions can matter, but the operator's convenient account must not define everything the overseer is able to examine.

Simulation adds another possibility. Before permitting an interaction, an institution could explore what might follow under stated assumptions. Different assumptions or models may reveal different consequences. That disagreement can help identify where further inquiry is useful. A simulated outcome remains conditional on the representation and assumptions; it cannot establish that every important effect has been anticipated.

These forms of help leave distinct governing questions open:

Connecting evidence can strengthen a decision's grounds. Exploring consequences can clarify the choice. Carrying out an instruction can give the choice practical effect. Institutions need to know which contribution they are relying on and which judgment remains theirs to justify.

This is the purpose of computational governance as used in this book: suitable computation makes governing work more observable, testable, executable, and capable of learning, while retaining legitimate authority and meaningful challenge. It includes conventional software as well as AI. Its usefulness is determined by what the whole arrangement can responsibly accomplish.

The opportunity extends across institutions. A developer may hold evidence, an operator may control a use, and an independent evaluator may question an earlier finding. Suitable computational methods could help connect that question to the permissions it bears on without requiring unrestricted pooling of records. Whether the actors can understand and respond together remains something to demonstrate. Shared computation would not make their mandates interchangeable.

Methods within a mandate

A method supplies capability, not permission. An institution may possess a tool able to inspect sensitive records while lacking authority to use it for a particular inquiry. It may have authority to investigate but need a narrower method to protect information unrelated to the question. Technical availability cannot determine the boundaries of access.

An authorized request can make the intended work clear. In the hypothetical laboratory, a reviewer might ask for a comparison between a proposed change and the conditions supporting an earlier evaluation. The request should identify the question and limits of the inquiry. It would not give the assistant an unrestricted mandate to collect information or alter the experiment.

A reusable method can help perform that comparison consistently. It might preserve the relevant versions, connect claims to records, and make unresolved differences visible. Reuse can conserve effort and support dependable treatment. The method still needs to remain suitable when the activity changes; repeating a comparison that no longer captures the important relationship would preserve consistency at the expense of understanding.

Some work may be delegated to a person or an AI system within defined permissions. The delegate could conduct a bounded inquiry, prepare a reasoned recommendation, or execute an instruction already authorized. Responsibility should distinguish who permitted the task, who carries it out, and who examines the result. A delegate's competence does not enlarge its remit, and its work is not automatically independent scrutiny.

Assistance need not stop at advice. Suppose an institution has legitimately decided to narrow a system's access. Suitable software could apply the restriction and help check whether it took effect. The institution would still need evidence of changed activity and a way to question it beyond the executing tool's own account. Confirmation that the instruction was dispatched or executed would not establish its effect.

An event-triggered check can make a method responsive to a consequential change. A new permission, changed interaction, or credible concern might require examination before the affected activity proceeds. The check needs coverage of the relevant routes, a defined effect, and a justified response when it is missing or fails. If an action can bypass it, calling it mandatory supplies little assurance. Not every change requires the same response.

These are ways of carrying out governing commitments, rather than a universal model to install. Public bodies, firms, and communities have different powers and obligations. A method useful within a company cannot give an external overseer powers it does not possess. Sharing a technical facility does not make the participants' mandates interchangeable.

We call the foundational commitments to purpose, authority, protected interests, responsibilities, and the conditions for challenging or changing decisions and governing arrangements the governance core. Adaptable mechanisms are the methods used to carry them out. A revised method may fulfill an existing duty more effectively. It may also change something substantive. Expanding sensitive-data access or removing an opportunity to challenge a decision requires scrutiny appropriate to that effect, even when the change is described as maintenance.

The methods themselves can become objects of deliberate development. An institution could compare how different arrangements discover a consequential change, preserve its context, and connect it to a timely response. Computation may assist that inquiry as well as perform the governing work. Its findings can support a proposal to revise an arrangement, retain what works, or seek further evidence; they do not authorize the revision. Chapter 11 examines how the institution's ability to learn can itself improve.

What would show that the help is useful

The strongest objection is that AI-assisted scrutiny may reproduce the same omissions while making them harder to recognize. If a reporting tool and a reviewing tool rely on the same incomplete material, agreement between them could reinforce a mistaken account. Greater fluency may make the account more persuasive without making it more dependable.

This objection limits what institutions should claim. Assistance needs evidence of usefulness for the particular work entrusted to it. A tool that reliably locates a relevant document has not thereby demonstrated reliable interpretation of its significance. A method tested on separate decisions has not necessarily demonstrated that it can examine their interactions. The task and conditions must remain attached to the claim.

Evaluation should compare the assistance with relevant alternatives. What did the existing method reveal or miss? Does the proposed method improve the grounds of judgment, preserve important context, and make uncertainty easier to examine? Errors and omissions matter alongside successful findings. Tests should include circumstances that could expose weaknesses, with conclusions limited to what was examined.

Developing the capacity to govern calls for testing whole arrangements as well as individual tools. A simulation or a test using synthetic records could examine whether a changed condition reaches a reviewer, whether that reviewer can pursue the question, and whether an authorized restriction covers the relevant activity. Such testing leaves transfer to actual operation unproven. Where consequential exposure is involved, the test itself needs justified permission and effective bounds; a promise of learning cannot supply them.

Demonstrated capability must also become usable institutional capacity. Can the responsible body maintain the method, interpret its findings, question its categories, and replace it? Accepted review practices, training, and contracts may need to change before the help can affect a decision. A correct finding that nobody is authorized or equipped to use remains a finding awaiting a governing response.

Evaluation must include that adoption burden and the attention consumed by questionable alerts. A method should earn reliance within the institution that will use it, with independent inquiry able to test its omissions.

We value public accountability over administrative efficiency.

Administrative efficiency is a genuine good. Less repetitive preparation can leave time for inquiry; shorter delays can make help more accessible. Where reliable assistance reduces unnecessary work, retaining the old method may itself impose avoidable burdens. The priority asks whether those gains support an institution whose decisions can be understood, challenged, and corrected.

More attention to the important question

A useful saving can make careful scrutiny practical. Its value depends on what happens to the effort released. Repairing the assistant's errors may divert reviewers from consequential questions, or missing work may fall to the people affected. Those burdens belong in the assessment of whether the arrangement has improved. The opportunity is more usable attention for consequential questions.

Assess the total burden. Preparation, integration, protection of information, checking, maintenance, and review all require resources. A developer might prepare a submission more quickly while an oversight body spends longer verifying it. An automated explanation might save institutional effort while making an affected person work harder to identify a mistaken premise. Who gains time and who loses it belongs in the judgment of improvement.

Fairness and transparency need similarly specific examination. A tool might help compare treatment across groups or make reasons more intelligible. Those contributions can be valuable where demonstrated. They do not decide which differences are justified or establish that important experiences were included. Chapter 4's concern about a summary treating an infrequent objection as minor applies here: numerical prominence cannot settle the significance of a burden.

Warranted confidence should follow the evidence of dependable conduct and meaningful correction. A persuasive explanation or increased acceptance is insufficient. Scrutiny may expose a deficiency and reduce confidence; that can be a useful result. Those evaluating the assistance must be able to question what counts as success, including costs or exclusions omitted from the preferred measure.

Independent inquiry remains necessary. Reviewers need access, competence, resources, and freedom to pursue an unwelcome question. Using a different model may be helpful without supplying those conditions. A person receiving recommendations needs time and practical ability to examine them. Where the available review arrangements cannot adequately check a consequential contribution, the institution should limit reliance rather than treat another review as an automatic cure.

Demonstrated usefulness does not require perfect performance. It requires grounds proportionate to the task and stakes, with knowledge limits affecting what is entrusted to the assistance. Some activity may need narrower permissions or simpler interactions even after computational capacity improves. Constraints and assistance can work together; increasing capacity does not justify every possible expansion.

More capacity and identifiable responsibility

The remaining question is how assistance changes the exercise of power. An institution could gain analytical capability while making its decisions harder for others to question. That would leave an important deficiency even if its work became more accurate or faster. Collective governing capacity includes the ability of independent reviewers and affected people to challenge the institution's account.

Consequential responsibility must therefore remain identifiable. Someone must be answerable for the purpose assigned to the tool, the permissions granted, and the decision to rely on its work. Responsibility for a particular result may be distributed across roles, but the connections must function. An affected person should not have to resolve the institution's internal divisions to obtain examination of a material concern.

A human signature alone cannot establish that arrangement. The reviewer needs relevant evidence, competence, time, and the ability to alter the course of the decision. Where the recommendation is effectively impossible to question, formally retaining a person at the end preserves little governing judgment. Repair may require better help, a different allocation of work, or limits on the activity being approved.

Useful help can also support people outside the institution. An intelligible account of a decision may help someone identify which premise to question or where to seek correction. Such assistance should preserve material differences and relevant uncertainty rather than translate every objection into the institution's preferred terms. People should be able to obtain help without first becoming experts in the system affecting them.

Dependence deserves attention as part of reliance. Convenient integration may be valuable, yet an institution needs to consider what happens when its governing tool fails or must be replaced. Can essential scrutiny continue? Can relevant records and reasons still be examined? Preserving that capacity may require resources and reduce some immediate convenience. The practical ability to change course will matter beyond this chapter.

The same awareness and scope obligations apply to the tools used for governance. An advice-only system can shape consequential decisions through what it includes, ranks, or omits. Institutions should understand that influence sufficiently for their responsibilities and preserve ways to discover material errors. Assistance should not gain an unrestricted role simply because another person formally authorizes the final output.

Return once more to the hypothetical laboratory. The assisted reviewer has a possible connection to examine, supporting records to check, and an unresolved question about an earlier condition. This is a useful contribution if it is reliable enough for that task and reaches someone able to pursue it. Whether the activity should continue still depends on the evidence, protected interests, legitimate authority, and the means to constrain its consequences.

We should expect institutions to develop governing capacity deliberately, demonstrate what computational help makes possible, and identify who remains answerable for its use. This shift requires investment in the means of inquiry and effective response alongside investment in the technologies being governed. People need not perform every operation unaided to preserve legitimate judgment. Better governing capacity can bring a consequential choice within reach; the next question is what response is justified now, including restraint or refusal.