Degrees of Maybe: Interface Language for Systems That Cannot Be Certain

A triage tool flags a benefit application for extra review. A high-street retailer's recommender surfaces a winter coat. A district council's assistant sorts a repair request into the wrong queue. In each instance a probability has been computed, weighed and acted upon, and in each instance the screen has told the user something that sounds like a fact. This is the quiet structural problem of the current product cycle: the machinery underneath our interfaces has become statistical, while the vocabulary on top of it remains administrative.

The flat tone of the certain interface

Digital services inherited their manners from deterministic software. A form either accepted a postcode or rejected it. A balance was a number. A status was one of four words. That inheritance made sense when the system's internal state was knowable and the interface's job was simply to report it.

Probabilistic components break that contract without changing the surface. A confidence score of 0.81 is rendered as a green tick. A classification the model gets right seven times in ten is presented with the same visual weight as a verified address. Layout, type hierarchy and colour all transmit certainty, and users, reasonably, read polish as precision. The interface is not lying, exactly, but it is speaking in a register it has not earned.

Calibration is a design problem, not a modelling one

Model teams are measured on accuracy, recall, or some composite of the two. Design teams are, or ought to be, measured on something adjacent but distinct: whether the trust a person places in an output matches the reliability of that output. Miscalibration runs in both directions, and both are expensive.

Under-trust produces what might be called redundant labour. A user who suspects the automation is unreliable will re-enter the same details, ring the contact centre, or abandon the digital route entirely. For a council or a department, that is a channel-shift cost with a queue attached. Over-trust is quieter and more damaging. When a person accepts a wrong classification without checking, the error does not stay local; it travels through case notes, eligibility decisions and, in commercial settings, purchases and refunds.

The stakes differ across sectors but the pattern does not. In public services, a mis-flagged application can delay a payment that someone was counting on. In retail, a badly calibrated recommendation is merely irritating until it is not — until it is a sizing suggestion attached to a return policy, or a credit offer dressed as a helpful hint.

Three registers of doubt

Not every output deserves the same treatment. A useful starting point is to sort probabilistic surfaces into three registers, chosen by the cost of being wrong rather than by the sophistication of the model.

Ambient. No numbers at all. The interface hedges in plain language — "based on the details we have", "a suggested answer" — and places a visible route to check the reasoning within the same view. This suits low-stakes sorting where the user's own judgement is the final arbiter.

Banded. Where a decision has consequences, the output should be expressed as a range or a ranked set rather than a point. "Most likely", followed by the strongest alternative, is often more honest than a decimal. Bands also survive model updates better than percentages, which invite a spurious precision the underlying data rarely supports.

Evidential. For decisions that affect money, eligibility or safety, the interface should show what the model used, what it could not see, and what would change the answer. "We cannot confirm this from the photograph you provided" is a sentence worth designing for. It converts an opaque judgement into a task the user can act on.

Error margins people can act on

Percentages are a poor delivery mechanism for uncertainty in most consumer and citizen contexts. Evidence from UK usability testing consistently shows that confidence figures are read as scores — 80% becomes a grade, not a probability. Bands with a tested legend travel better: likely, roughly even, unclear. The critical discipline is consistency. A single scale, used across every surface in a service, taught once and never redefined, does more for calibration than any amount of explanatory microcopy.

Where the uncertainty has a shape, show the shape. A model that is confident about the category of a problem and uncertain about its urgency should say so separately, rather than collapsing both into one hedge.

Fallback is a state, not an error

Most teams still treat the uncertain path as exception handling: the modal that appears when the model shrugs. That framing guarantees it will be under-designed, under-tested and slower than the confident path. The opposite is required. The "we are not sure" route should be as short as, or shorter than, the automated one, and it should be reachable from the same screen without a hunt through help pages.

This matters particularly where authority has been delegated — an adviser acting for a claimant, a carer managing a relative's account, a member of staff operating on a customer's behalf. Those intermediaries need the confidence signal as much as the end user, and often in a more compressed form. If the public interface says "likely" and the staff interface says "matched", the two systems are already disagreeing.

Designing for calibration rather than comprehension

The usual research question — did users understand the output? — is too weak. The better question is whether their behaviour predicts the system's reliability. Ask participants to state, before acting, how often they expect the system to be right; then compare that estimate with the observed rate. Track overrides, escalations and repeated submissions as calibration signals in production, not merely as error metrics.

Governance is moving in the same direction. The ICO's guidance on explaining AI-assisted decisions, the Algorithmic Transparency Recording Standard used across UK public bodies, and the outcomes-focused expectations of the FCA's Consumer Duty all converge on a single demand: that affected people be able to form a reasonable view of how a decision was reached. None of them prescribe interface patterns, which is precisely why design has room to lead rather than comply.

Interfaces are confidence statements whether or not we intend them to be. The question is whether the statement is designed or inherited.