What Are We Assuming?

The grounds for staying

A decision to keep relying on an arrangement rests on beliefs about what will continue to work. A supplier will remain available. Relevant evidence can still be examined. A reviewer can recognize an important change. The previous chapter asked whether people and institutions retain practical choices when they depend on AI. Those choices become harder to exercise if nobody notices that the grounds for continued reliance have changed.

An assumption need not be careless or unsupported. It may summarize experience, preserve the result of a sound evaluation, or allow work to proceed without repeating an inquiry already completed. Institutions need such expectations. People need dependable arrangements rather than a fresh negotiation whenever they seek help.

The difficulty is that an expectation can survive the conditions that justified it. A review method might have been adequate for an earlier permission and become inadequate after the activity changes. The original judgment could remain reasonable in its context while no longer supporting the present decision. Recognizing this difference allows examination without turning every reconsideration into an accusation against those who came before.

Questioning the belief does not necessarily release the institution from the arrangement that sustains it. A review may remain familiar because people are trained to perform it, contracts require its outputs, and other decisions depend on its timetable. Even after its limits become clear, an alternative may have no recognized place in the work. The assumption concerns what is believed to be adequate; the institutional question is what keeps that belief in use.

Consider a hypothetical authority that assumes a scheduled review will remain sufficient until the next review date. Earlier stable permissions and dependable records could support that expectation. Suppose the system then gains permission for direct actions whose effects can occur before that date. Staff still prepare the accepted report, the budget provides for the scheduled meeting, and completed reviews remain the evidence that a duty was discharged. Each arrangement makes the others easier to retain. Bringing review forward now requires changes that repeating the familiar process does not.

This would warrant examining the reinforcement as well as the expectation. Delay alone would not establish path dependence: a lack of funds, a binding duty, or legitimate disagreement could independently explain it. The question is whether earlier choices have made the established method easier to reproduce while narrowing practicable alternatives. The Governance Assumptions Register introduced next records expectations, their grounds, reasons to reconsider, and who should notice. Institutional reinforcement needs a connected inquiry, developed further in Chapter 11.

We should make consequential assumptions explicit, especially where being wrong could remove a protection or close a practical choice. That does not require listing every institutional belief or presuming inherited arrangements mistaken.

A useful distinction comes first. An institution may be committed to protecting people's access to a service while assuming that a particular review will detect an unjustified exclusion. The commitment and the assumption serve different purposes. Evidence that the review misses something gives reason to examine the method. It does not make the protected interest obsolete or authorize the institution to abandon its duty.

We will begin with four questions that make assumptions examinable, then consider several expectations worth questioning in context. The final movements connect a changed assumption to the decision it supports and ask who can recognize the change in time.

Four questions worth keeping open

The Governance Assumptions Register proposed in this book is a simple discipline for keeping expectations open to evidence. Its public form has four questions:

What are we assuming?

Why do we believe it?

What would make us reconsider it?

Who will notice when that happens?

The first question asks for a proposition someone could examine. “Human oversight” names an arrangement. “The assigned reviewer can identify a consequential error before approving this action” states something the arrangement is expected to accomplish. Making that expectation explicit helps reveal which evidence and practical conditions matter.

The second question respects the reasons for relying on it. A relevant evaluation, professional experience, demonstrated access to records, or a dependable response route may support the belief. The grounds should fit the claim. Evidence that a reviewer can understand separate decisions may leave open whether they can examine interactions among them. A sound reason for a narrower belief should not quietly become assurance for a broader one.

The third question gives the expectation a way to encounter correction. Changed permissions, a credible complaint, an independent finding, or a loss of necessary evidence could matter. So could a useful new method that makes an earlier burden avoidable. Reconsideration is not reserved for failure. An assumption about what cannot reasonably be done may deserve attention when new capacity becomes available.

The fourth question concerns whether that attention is practical. Someone must be able to receive the relevant evidence, recognize its significance, and bring it to an appropriate decision-maker. An entry that ends with a name supplies little assurance if the person lacks access, time, or a route to response. We will return to this question after examining what may need to be noticed.

The record can be brief. A complete assumption, its grounds, a reason to look again, and a responsible function may be enough. Where authorization depends on it, add which permission would need reconsideration. The purpose is to make a consequential relationship usable in judgment. A perfectly completed form would not establish that anyone can recognize the change it describes.

Appendix C provides a short form and four worked illustrations spanning operational reliance and institutional improvement. They show how to use the questions rather than catalogue every expectation discussed here.

There is a serious objection to doing this too extensively. Examining every premise can consume the attention needed for urgent work. People may face delay while an institution repeatedly reopens a question already settled on adequate grounds. Concentrate on assumptions whose failure would materially affect a decision, protection, or capacity to act. Preserve sound expectations, and give reasons for demanding another examination.

Reasonable expectations under different conditions

Consider the expectation that reliable reporting requires extensive manual preparation. It can have a sensible origin. Records may be scattered, sensitive information may need careful handling, and a professional may have to judge what an account means. A demanding preparation process can reflect necessary work rather than an attachment to inconvenience.

Reliable assistance or reusable evidence could change part of that work. If a method can assemble relevant records dependably, a previously impractical question might become reasonable to ask. The assumption to reconsider is what useful evidence can be obtained at a proportionate burden. It is not that every report should now be produced automatically or that greater volume would improve scrutiny.

We value useful evidence over comprehensive reporting.

Comprehensive reporting preserves context, exposes omissions, and makes comparison possible. These benefits matter when a reporting party would prefer a convenient subset. The priority asks what the account enables someone to understand or question. Reliable assistance might justify asking for more relevant evidence, while removing repetition that consumes attention without helping the decision.

The full burden needs examination. Less preparation for the sender could mean more checking for the recipient. Integration, confidentiality, maintenance, and correction remain work. A tool that makes a report more convincing has not established that its evidence is more reliable. The reporting owner and recipient should examine what actually becomes practicable, with appropriate independent scrutiny, before declaring the old cost assumption obsolete.

A second expectation is that human approval establishes meaningful oversight. Its reasonable origin is the capacity of a responsible person to consider context, ask questions, and withhold permission. Human judgment can matter where a fixed check cannot capture what is at stake. Assigning that role can make responsibility identifiable and create a genuine opportunity to alter the result.

But the signature depends on conditions. The reviewer needs relevant evidence, competence, time, sufficient independence, and practical influence over the decision. If these are absent, adding another approval may repeat the same gap. The institution assigning the duty should ask what the person can actually understand and change. Repair should preserve needed judgment and responsibility while making the duty possible to fulfill.

That inquiry applies equally to AI-assisted review. Assistance might make a demanding comparison manageable, or it might reproduce the same omission in a more fluent account. Agreement between a reporting system and a reviewing system does not establish an independent basis for confidence. The question is whether the combined arrangement supports the judgment being entrusted to the person, including a way to pursue unwelcome evidence.

A third expectation is that a comprehensive, convincing explanation supplies sufficient understanding. Breadth and explanation deserve respect. They can connect events, reveal an omission, and make a technical matter intelligible to someone exercising a different responsibility. An institution should not dismiss an account merely because it is well written or prepared with assistance.

The issue is what supports its consequential claim. If an explanation attributes an action to a particular cause, what evidence allows that attribution to be checked? If the account says a safeguard remains effective, what activity and conditions does the evidence cover? A plausible account can assist inquiry while leaving those questions unresolved. Its usefulness depends partly on making the remaining limits visible.

An independent finding or affected person's experience may prompt another look even when the official account remains unchanged. The contribution need not prove the whole account false. It may reveal that the account answers a narrower question than the decision requires. Appropriate access, follow-up, and investigation can then help establish what should still be relied on.

The following comparison brings together these three expectations. They are contextual assumptions to examine, not beliefs attributed to all institutions or conclusions that the assumptions are false.

Preparation effort, a responsible person's approval, and an intelligible explanation can each be valuable. The comparison asks when they support the governing achievement expected of them. Making the expectation explicit lets us examine the conditions of reliance without presuming the assumption mistaken.

Review can confirm what works

An examined assumption may remain sound. It may also prove dependable within a narrower scope than first understood. Both findings can preserve a useful arrangement. Learning has a purpose when it clarifies the grounds of reliance, even when the result is to keep a protection or method in place.

Which permission depends on it

A further expectation brings the question directly to continued operation: previously authorized activity remains adequately overseen after a consequential change. Earlier evaluation may have supplied good grounds for the original permissions. The reason to reconsider is that tools, interactions, access, or governing methods can change what those grounds support, even without a new model.

Return to the hypothetical public service extended in the previous chapter. Suppose its permission to make certain changes directly depends on a review method that lets officials examine those actions in time to respond.

Now imagine a change in how the system's actions reach that review. Officials cannot yet establish whether the method still covers the directly authorized actions. This is an added hypothetical condition, not an established oversight failure or account of an actual service.

The proposition becomes specific: the review method still supplies adequate grounds for overseeing the directly authorized changes. Its reasonable basis is the earlier evaluation within the earlier arrangement. The reason to reconsider is the changed pathway and uncertainty about coverage. The affected decision is continued permission for those changes, rather than every use of AI in the institution.

Making the connection specific helps distinguish a concern from an indiscriminate demand to start again. Advice that remains adequately examinable might still be justified. Some directly authorized activity might remain covered too. Continuing those activities would require evidence that they can remain adequately overseen and constrained while the affected activity is restricted. The institution should not assume this simply because keeping part of the service running would be convenient.

Nor should it wait automatically for a demonstrated injury. Where permission depended on a safeguard remaining examinable, a material loss of that evidence can undermine the basis for permission before harm is proved. The inquiry concerns which grounds remain available now. An earlier approval records a judgment made under earlier conditions; it cannot supply evidence of present coverage that the institution lacks.

This does not require one response to every missing record or update. Significance depends on the activity, consequences, evidence needed, and constraints that remain effective. Appropriate responses could include restoring evidence, adding review, narrowing a permission, reverting a change, or pausing affected activity. The competent authority must justify the response, including its costs and the effects on people who need the service.

We value learning through action over advance planning.

Advance planning helps identify dependencies and anticipate what a change could invalidate. It can prevent exposure that later inquiry would not repair. Learning through action asks that preparation connect to justified opportunities to examine what actually holds. Investigating coverage, testing a method with synthetic records, or enforcing a narrower permission can be action. None requires exposing people to an unjustified experiment.

A test would still have limits. Showing that a revised method covers a simulated pathway would not establish that it works under every operating condition. The claim should retain the conditions examined and the uncertainty left open. Where the evidence remains insufficient, that limit must affect reliance rather than appear only in a note attached to unchanged authorization.

Discovery also has limits of authority. A reviewer who identifies the questioned assumption may be entitled to investigate or raise it without being entitled to rewrite the rule. A provider proposing a technical repair does not thereby acquire power to expand access to private information or remove a route to challenge. The protected commitment remains binding while the means of fulfilling it are reconsidered through appropriate authority.

This protects against another misuse of the register. Someone may describe a safeguard as an outdated assumption because it obstructs a preferred activity. The relevant inquiry must distinguish a claim about how protection is achieved from the legitimacy of the protection itself. Changed feasibility can support a different method. It cannot alone settle who may decide or whose interests may be overridden.

Someone able to notice

The remaining question is who recognizes that the grounds have changed. In the hypothetical service, responsibility could involve the operating team identifying a changed pathway and an oversight function examining its significance. These roles need a working connection. A technically accurate finding that never reaches anyone able to reconsider permission would leave the governing task unfinished.

Noticing requires access to more than reassuring summaries. The responsible function needs evidence appropriate to its duty, competence to interpret it, and resources to pursue a consequential question. It also needs routes beyond the operator's account. Workers, independent evaluators, and people affected may encounter something the established method misses. Protected reporting can make those contributions possible without requiring every reporter to become publicly identifiable.

No individual can foresee every surprise. Specifying a reason for reconsideration helps focus attention, but should not exclude evidence that arrives in an unexpected form. A complaint can matter even when it does not fit an existing field. The institution should carry its share of inquiry rather than require the person raising the concern to prove an entire failure unaided.

The review owner needs a route for consequential changes to receive attention between scheduled reviews, with scope and burdens justified by the affected permission. Changes to tools or permissions could be notified to that owner, alongside routes for complaints and unexpected findings. The duty needs staff time, usable evidence, and authority to be fulfilled.

Computational assistance could help connect a changed condition to earlier reasons and permissions. That contribution would need evidence of reliability for the task and scrutiny of what it omits. The means used to notice change have assumptions of their own. Their usefulness must remain open to independent inquiry, rather than become the only route through which their adequacy can be questioned.

We should expect consequential reliance to have grounds that can be examined and a person or function equipped to recognize when those grounds need another look. This gives dependable arrangements a way to remain dependable under changed conditions. It can preserve useful assistance, protect a necessary limit, or reveal where present permission is no longer supported.

The expectation also gives people a more reasonable relationship with institutions. They need not reconstruct every governing assumption to obtain help. A material concern should be able to reach competent attention, with someone responsible for what follows. Keeping that capacity effective as circumstances change brings us to the next question: how does governance learn?