How Does Governance Learn?

Beyond correcting one decision

A material concern should be able to reach competent attention, with someone responsible for what follows. The previous chapter ended with that expectation. Fulfilling it once would be worthwhile. Keeping it effective as circumstances change requires an institution to learn about the work through which it governs.

Return to the hypothetical public service discussed in earlier chapters. Someone had asked whether an account of faster completion included people unable to begin or finish. That question supplied grounds for inquiry, without establishing exclusion or its cause. Now add a further hypothetical condition: imagine inquiry establishes a mistaken decision that the competent authority can correct.

Correction matters to the person affected. They should not have to wait for institutional redesign before receiving help already justified within the authority's remit. But the finding also raises a question about future decisions. Did the institution's evidence requirements leave out relevant experience and, if so, did that omission contribute to the mistaken decision or leave similar matters unrecognized?

If inquiry identifies such an omission, the institution could revise what it asks for so that its examination includes people absent from the completion measure. It could then establish a dependable way for later complaints and independent findings to prompt reconsideration of those requirements. These are possible developments within the illustration, not achieved remedies. Each would need authority, resources, and evidence of usefulness.

These possible responses address different problems. Correcting the decision addresses one matter. Improving the requirement changes what future review can examine. Improving how requirements are reconsidered gives unfamiliar concerns a way to change the questions themselves. Learning reaches beyond an answer to the arrangements that make a better answer possible.

We will consider what should remain dependable, where learning needs to reach, and how its methods can improve. The final movements ask who may authorize a change and what would establish that it made governance better.

Dependable commitments and adaptable methods

People need institutions they can rely on. An essential service should not become a fresh negotiation whenever a new tool appears. Stable responsibilities preserve experience, make treatment predictable, and constrain the discretion of those holding power. Changing less can sometimes be the responsible choice.

Yet an unchanged method can cease to fulfill a dependable commitment. A review designed around advice may not examine direct action. A reporting requirement may preserve a useful comparison while excluding experience the institution now needs. Continuity of method and continuity of protection are related, but one does not establish the other.

Even after recognizing that failure, the institution may be unable to repair it. Earlier choices can organize later investments around one method until alternatives become difficult to pursue. This is the concern about path dependence introduced earlier. Its significance lies in how an arrangement reinforces its continued use, rather than in its age alone.¹

The governance core introduced in Chapter 6 sets out the foundational commitments of the institution. It concerns what the institution is entrusted to do, who may decide, and what must remain protected.

Adaptable mechanisms are the ways those commitments are carried out. They include reporting, evidence access, analytical methods, delegation, interventions, coordination, and review. People, procedures, conventional software, and AI can contribute. A different method of examining access might fulfill the same responsibility more effectively, provided its contribution can be established.

Core and mechanisms are two views of the same arrangement. They do not create another institutional hierarchy. Foundational deliberation uses methods too: how proposals are examined and whose experience enters reconsideration can matter when a governing purpose is revised. Adaptability therefore extends beyond routine administration.

Neither part is exempt from scrutiny. A foundational commitment may need legitimate reconsideration. An operating method may carry a protection that cannot be removed casually. Expanding access to sensitive records or eliminating a route to challenge changes something substantive, even when described as a technical adjustment. The authority required follows what changes and whom it affects.

In the hypothetical service, an inadequate account of access would give reason to examine the evidence requirement. It would not authorize abandonment of access protections. Nor would a proposed repair acquire legitimacy merely because it could be implemented quickly. Dependable foundations and adaptable ways of working need to support one another through justified change.

Suppose the service trained reviewers to assess completed applications, purchased reporting software around that task, and agreed that its supplier would deliver the resulting totals. Over time, expertise, contracts, and evidence requirements become fitted to one another. A request to examine people unable to begin now needs different evidence, skills to interpret it, and permission to obtain it. Even willing reviewers cannot produce these conditions by recognizing the omission.

Recognition can also disturb what the institution rewards. If completion of the accepted review demonstrates competent work, questioning its adequacy may expose earlier omissions or unsettle professional roles. A familiar method can appear easier to defend while an alternative carries the burden of proving both reliability and authority. These are possibilities within the illustration. An actual diagnosis would need evidence of that history and reinforcement. Resource scarcity, political opposition, or sound reasons for continuity could explain difficulty changing course without that mechanism, or could interact with it. Each needs examination on its own grounds.

Where learning needs to reach

To learn from the mistaken decision, the institution needs to examine more than its final answer. What was available before judgment? How was it assessed? What did the arrangement produce? What happened to people because of it? These questions distinguish inputs, processes, outputs, and outcomes.

From the overseer's perspective, inputs include complaints, test evidence, expertise, resources, and affected people's experience. Authority also matters: a reviewer needs powers appropriate to the responsibility. An account can be accurate yet inadequate if relevant experience never enters it or nobody has time to pursue the question it raises.

Processes concern how that material is examined, disagreements handled, decisions authorized, and responses implemented. Could the reviewer ask a follow-up question? Could an unexpected concern reach someone able to act? A richer submission would accomplish little if the process excluded every contribution that did not fit its established categories.

Outputs are what the arrangement produces: an approval, a restriction, a revised requirement, or an explanation. Outcomes concern what happens to people, institutions, and relevant conditions. A corrected instruction is an output; restored access would be an outcome. Evidence that the instruction was issued leaves open whether it changed what the person could actually obtain.

The perspective matters. The service's report is its output and may be an overseer's input. The labels help identify which relationship is being examined. Their purpose is to keep improvement from stopping at the artifact most convenient to count.

Learning might therefore require a different question, additional expertise, or resources for inquiry rather than a more sophisticated analytical tool. It may require incentives that make unwelcome findings usable. If candid uncertainty is treated as incompetence, asking for fuller accounts could encourage more confident presentation without improving understanding.

Evidence requirements also impose burdens. A new request might reveal an omitted experience while requiring people already struggling to obtain help to document every difficulty. The institution should examine other reliable ways to obtain what it needs, protect sensitive information, and carry its share of investigation. More information is useful when it strengthens judgment at a proportionate total burden.

Producing evidence, checking it, and communicating it remain different work. AI could help assemble records or make an account intelligible, but a fluent description cannot supply an investigation that never occurred. Learning includes asking whether the evidence was generated and adequately examined, as well as whether its presentation improved.

That brings evaluation itself into view. Suppose the service assesses every revision through completion speed. A change could look successful while leaving access unexamined. The institution would need to reconsider how it identifies improvement, including whose experience and which protections its preferred measure leaves outside the judgment.

Improving the ability to improve

The public-service illustration now lets us distinguish different objects of improvement. These are possibilities to examine, rather than stages every institution must follow.

The last row opens a further question: can the route for future concerns recognize its own omissions? An institution might receive complaints efficiently while overlooking people unable to complain. It might revise requirements promptly while allowing only the reporting party to propose revisions. Improving the route would require examining those conditions, not merely processing more contributions through it.

Iterative governance improvement means repeatedly improving what governance asks for, how information is produced and checked, how decisions are made and implemented, and what happens as a result. It reaches the work we have been examining, including its inputs and consequences.

Recursive governance improvement, in the stronger sense used in this book, means improving the arrangements that make those improvements possible: how new questions are identified, reporting requirements are revised, evidence is evaluated, affected people participate, and institutions recognize their own failures.

The categories overlap: repeated improvement can change the improvement process itself. Revising a report can improve a decision; changing how excluded concerns shape future reporting can also improve how the institution learns.

Institutional learning also has established antecedents. The OECD’s framework for anticipatory governance, for example, brings adaptive regulation together with foresight, participation, and international cooperation. This book develops a public argument about the capacity and accountability such learning requires.²

In the service, this could mean giving independent findings and concerns outside the usual form a route into reconsideration. It could mean enabling reviewers to develop alternatives and bring them to the authority entitled to decide. It could mean changing the evaluation method when an apparently successful reform omits a consequential burden. Each proposal needs scrutiny of its own effects.

Suitable computation could assist this inquiry. A tool might help compare past decisions, identify recurring omissions, or investigate how several changes interact. Its usefulness would need demonstration for that work. The institution must also be able to examine what the tool misses; using the same method to propose and judge every improvement could preserve its blind spots.

The Governance Labs proposed in Chapter 8 could contribute by investigating alternatives across institutions. Their findings could inform a responsible owner's decision to retain or revise a method. Research would not supply the authority to adopt it. The learning arrangement includes the connection between inquiry and legitimate decision, alongside the quality of the inquiry itself.

Learning need not wait for harm. A test, a near miss, a useful new capability, or better evidence could reveal a worthwhile revision. Testing with synthetic records might show that an unfamiliar concern never reaches a reviewer. That finding would concern the conditions examined, leaving performance in actual operation to be established.

We value learning through action over advance planning.

Advance planning can prevent an exposure that no later correction would undo. The priority asks it to lead into justified opportunities to learn where those exist. Investigation, simulation, enforcing a limit, or retaining a protection after examination can be action. A promise of learning does not authorize experimentation on unwilling people or dissolve rights and binding duties.

Inquiry should concentrate on consequential deficiencies and retain the reasons for methods whose grounds remain sound. That memory can prevent repeated pursuit of an alternative already rejected for a reason that still matters.

Who may change what and when

Improvement requires an authorized owner able to connect inquiry with implementation. In the service, the duty to protect access and the immediate correction already warranted should continue while the evidence method is reconsidered. Feasible alternatives might include a narrower new inquiry, a revised reporting request, or assistance that helps reviewers examine missing experience. Their expected contribution and limits need comparison with the present method.

A bounded, authorized test could help establish whether an alternative answers the omitted question. Adopting it may then require staff time, training, resources, a changed supplier agreement, and acceptance of the new evidence within review. Someone must be responsible for securing those conditions or explaining why they cannot be secured. A successful demonstration would otherwise leave the old arrangement as the only usable one.

These are connected questions, not compulsory stages. As evidence becomes available, the owner should examine the method’s effects and burdens with independent scrutiny, retaining, revising, or reversing it where justified and feasible. The transition must preserve remedies and essential protection while making a better method practicable. Where that cannot yet be done, the permission relying on inadequate oversight needs a justified limit.

Different revisions warrant different decisions. Adjusting how an authorized analysis locates relevant records may fall within an existing remit. Expanding which private records it may inspect changes the scope of access. Removing an appeal changes protection against error or arbitrary treatment. A supplier's ability to install all three changes would not make its authority to do so equivalent.

The same applies when the core is reconsidered. Experience may reveal that a responsibility is assigned to an institution unable to discharge it, or that an existing allocation of powers leaves a consequential gap. That can justify proposing a foundational revision. It cannot give the proposing institution permission to enlarge its own powers. Legitimate participation, applicable obligations, and appropriate authorization remain necessary.

Timescales should follow these differences. A permitted operational adjustment may be made promptly. Changing a policy may require broader examination. Reconsidering authority or a fundamental protection may properly take longer still. Deliberation is part of responsible change where people need to understand and influence its terms.

Longer deliberation must not become an excuse to continue activity whose current permission lacks adequate grounds. If oversight has materially lost evidence on which authorization depended, the competent authority may need an interim response: restore evidence, add review, narrow the affected permission, revert a change, or pause as justified. Such a response can preserve protection while larger questions remain unresolved.

This is not a rule that every uncertainty demands suspension. The stakes, safeguards, and time available to prevent or correct a consequence matter. Governance need not revise itself as frequently as AI is updated. It needs methods current enough for the activity it permits, with stable protections and a response that can matter in time.

There is a substantial cost to making change answerable. Consultation, independent examination, and evaluation use resources. Frequent revisions can exhaust staff and make a service harder to navigate. These are reasons to justify the scope and pace of revision, compare alternatives, and simplify unnecessary work. They are also reasons to retain a sound arrangement rather than advertise change as an achievement in itself.

What would establish improvement

The remaining question is what would make a revision better. An institution should identify the deficiency addressed, compare the existing arrangement with feasible alternatives, and explain what evidence could change its judgment. It should preserve enough of the reasons and relevant conditions to make later comparison meaningful.

Outcomes matter, but the basis for judging them needs care. A useful test can supply evidence before a proposed method is entrusted with consequential work. An observed improvement after adoption may support continued reliance while leaving its cause uncertain. Changes in demand, resources, or permitted activity could also affect the result. State what the evidence supports rather than attribute every favorable change to the revision.

In the hypothetical service, a revised requirement might make missing experience easier to identify. That would be useful intermediate evidence. Establishing that access improved would require examination of what happened to people, including those still absent from the institution's usual account. Faster reporting or more completed reviews could not settle that question alone.

A better measure can reveal a worse picture

A more usable way to raise concerns could expose problems the old arrangement missed. Recorded concerns might rise even as the institution becomes better at hearing them. A fall might instead reflect a harder route to complain. Changing how we learn can change what we see; the direction of one indicator cannot settle whether governance improved.

Evaluation should therefore examine the routes through which evidence becomes available, as well as the findings received. Independent inquiry and affected people's experience must be able to challenge the account of success. Their contribution can reveal a missing consequence or dispute which burden is considered acceptable. Better measurement informs that choice without removing the need for legitimate judgment.

Distribution belongs in the comparison. A revised method could save a central institution time while shifting preparation and checking onto smaller partners. It could improve routine treatment while making an unusual circumstance harder to question. Ask what became better, for whom, at what cost, and with what uncertainty. An unrelated gain cannot compensate automatically for a missing essential protection.

We value lasting benefit over immediate gains.

Immediate help deserves weight: people needing a service now should not bear indefinite delay for an unspecified future advantage. Lasting benefit asks whether the improvement preserves dependable protection and future choices. That may justify investment in institutional memory, continued scrutiny, and the practical ability to change course, with the present burden explained.

Where feasible, an unsuccessful revision should be reversible. Retaining reasons and relevant evidence can help establish what to restore or change again. But restoring a method cannot undo every consequence produced while it operated. Lost opportunities may remain lost, and the loss of privacy caused by disclosure may be irreversible. Limits on recovery can justify caution before exposure, while duties to investigate and provide appropriate remedy continue afterward.

An evaluation may also conclude that the existing arrangement should remain. A protection retained on examinable grounds is a result of learning. So is declining a proposal whose advantage has not been established. Neither finding needs to be disguised as a successful update. The institution should be able to admit an incomplete answer and identify what remains justified while inquiry continues.

We should expect governance to improve its decisions, its questions, and the arrangements through which both can improve. The claim must remain attributable and open to challenge, with someone responsible for acting on what is learned. People should be able to obtain correction without redesigning the institution themselves, and benefit from learning that makes future concerns easier to recognize and address.

That expectation leaves power with accountable institutions rather than with an unquestionable improvement process. The next question is who can examine those institutions and their tools, challenge their definition of success, and make correction effective when the governors themselves are wrong.