Screens to Systems: Why Usability Is No Longer Just Interface Work
The views expressed are my own and do not represent any organization I am affiliated with.
Introduction
For decades, usability practice has focused on interfaces. We evaluate screens, workflows, labels, and controls because those are the visible points where users encounter a system. Confusing layouts, unclear feedback, poor error handling, and inconsistent navigation have been reliable indicators of deeper problems.
That model still holds for many systems. But an increasing number of tools now behave in ways that extend beyond what is visible on the screen. They adapt, infer, and respond based on context, history, or probabilistic reasoning. In these systems, the interface can be clear and compliant while the overall experience still feels unreliable or difficult to trust. This is not a failure of usability practice. It is a signal that the object of evaluation is shifting.
A Scenario Worth Considering
Imagine a clinical decision support tool used by nurses during patient intake. The interface is clean: vital signs go in, risk flags come out, color-coded for urgency. A usability evaluation finds no significant issues. Labels are clear. Workflows are logical. Error states are handled gracefully. The system passes every heuristic.
Yet over time, nurses begin double-checking its outputs against their own judgment. They document hedging language in their notes. Some stop mentioning the tool’s recommendations to physicians altogether. When interviewed, they struggle to articulate the problem. The system works. They just do not rely on it.
What happened? The tool uses an adaptive risk model. It adjusts thresholds based on hospital-wide patterns, patient history, and seasonal factors. Its outputs vary in ways that are clinically defensible but experientially inconsistent. A patient who flagged yellow last week might flag green today with identical vitals because the model’s context shifted.
The nurses are not irrational. They have learned that the system’s behavior is not stable enough to anchor their professional judgment. The interface told them nothing was wrong. The system’s behavior told them something else entirely.
This is the core challenge. The system is no longer just executing instructions. It is interpreting intent. And interpretation, by its nature, introduces variability that users must learn to navigate without guidance from the interface itself.
The Legacy Assumption in Usability
Most usability methods assume that system behavior is stable. A task performed today will behave the same way tomorrow. If a user learns how a system works, that understanding remains valid until the system is intentionally changed.
Heuristic evaluations, usability tests, and information architecture reviews are all built around this assumption. They examine whether users can form accurate mental models, execute tasks efficiently, and recover from errors within a predictable environment. The interface is treated as the primary mediator between user intent and system action.
This assumption has been reasonable for most of the history of digital systems. It remains reasonable for most systems in operation today. The challenge is recognizing when it no longer applies.
A Spectrum of Adaptive Behavior
Not all adaptive systems present the same usability challenges. It helps to distinguish among several categories.
Some systems personalize based on explicit preferences. A user sets their language, their notification schedule, and their default view. The system adapts, but the adaptation is transparent and user-controlled. Traditional usability methods handle these systems well.
Other systems infer from behavioral patterns. A search engine learns from click history. A content feed reorders based on engagement. A scheduling tool adjusts recommendations based on past acceptances. Here the adaptation is less visible. Users may notice that results feel tailored without understanding why or how to change it.
A third category generates probabilistic outputs. A diagnostic tool produces confidence scores. A recommendation engine surfaces options ranked by likelihood. A language model generates responses that vary with each query. In these systems, identical inputs may produce different outputs, and the system itself may not be able to explain why.
Each category creates different usability challenges. Explicit personalization requires good settings interfaces. Behavioral inference requires transparency and control mechanisms. Probabilistic outputs require calibration support and appropriate uncertainty communication. A usability practice that treats all adaptive systems identically will miss important distinctions.
Why Traditional Methods Strain
Usability specialists increasingly encounter situations where established methods produce incomplete explanations. A heuristic evaluation may find no significant interface violations, yet users report discomfort or hesitation. A usability test may show task completion alongside repeated verification behavior or reluctance to rely on system outputs.
These findings are often described informally. The system technically works, but it feels unpredictable. Users succeed, but they do not trust the result. Nothing is broken, yet something is wrong.
Traditional methods strain here because they are optimized to detect interface breakdowns, not behavioral uncertainty. They ask whether users can complete tasks, not whether users can calibrate their reliance on system outputs. They measure efficiency and error rates, not trust formation or confidence calibration. These methods were designed for a different class of problems. Recognizing that boundary is the first step toward extending them.
Trust as a Usability Concern
In traditional systems, trust is largely a matter of consistency and feedback. If a system behaves predictably and provides clear feedback about its state, users develop appropriate reliance. They learn what the system can and cannot do. They learn when to trust it and when to verify.
In adaptive systems, trust becomes more complex. Users must develop calibration: an accurate sense of when system outputs should be relied upon and when they should be questioned. This requires not just consistent behavior but legible behavior. Users need to understand, at least approximately, why the system responded as it did.
Calibration failures take two forms. Over-reliance occurs when users trust system outputs in situations where the system is unreliable. Under-reliance occurs when users discount system outputs that would have been valuable. Both are usability failures, but neither is visible in traditional interface evaluation.
Recovery matters too. In a traditional system, error recovery means helping users get back on track after a mistake. In an adaptive system, recovery also means helping users understand when the system’s inference was wrong and why. A user who cannot diagnose why the system produced an unexpected result cannot learn to calibrate their reliance appropriately.
The Specification Problem
Traditional usability evaluation assumes that correct system behavior can be specified. A button should do what its label says. A workflow should lead to the expected outcome. An error message should accurately describe what went wrong. Evaluation consists of checking whether actual behavior matches intended behavior.
In adaptive systems, intended behavior is often probabilistic or context-dependent. A recommendation system is not supposed to produce the same results every time. A risk model is supposed to adjust based on context. A generative system is supposed to vary its outputs.
This creates a specification problem. If correct behavior is inherently variable, what does usability evaluation check against? The answer cannot be a fixed specification. It must be something more like a behavioral envelope: a range of acceptable variation within which the system can be considered to be working as intended.
Defining a behavioral envelope requires collaboration across roles. Engineering teams understand the technical bounds of system behavior. Data science teams understand model performance characteristics and known failure modes. Product teams understand which variations are acceptable to users and which would constitute failures. Usability specialists bring the user perspective: can people perceive when the system is operating within bounds? Can they tell when something has gone wrong? Do they have recourse when it does?
Without a defined envelope, usability evaluation lacks a reference point. With one, evaluators can assess whether the system communicates its operating state, whether users develop accurate intuitions about reliability, and whether the boundaries of acceptable variation align with user expectations.
The New Unit of Analysis
In adaptive systems, the interface is no longer the only meaningful unit of analysis. The system’s behavior over time becomes equally important. Patterns of response, consistency across sessions, and the legibility of system reasoning all influence usability outcomes.
This shifts the evaluator’s focus from isolated interactions to interaction trajectories. The question is no longer only whether a user can complete a task, but whether they can develop a stable understanding of how the system behaves and when it should be relied upon.
Toward Extended Methods
What might extended usability methods look like? Several directions seem promising.
Longitudinal observation becomes more important. Single-session usability tests reveal whether users can complete tasks, but not whether they can develop appropriate reliance over time. Evaluating adaptive systems may require observing users across multiple sessions, tracking how their mental models evolve as they accumulate experience with system variability.
Consistency testing takes on new meaning. Rather than checking whether the interface is consistent, evaluators might check whether system behavior is consistent in ways that users can perceive and learn from. This might involve presenting users with scenarios designed to reveal the boundaries of system adaptation and assessing whether they can predict when behavior will change.
Mental model assessment expands. Traditional mental model research asks whether users understand the interface. Extended assessment asks whether users have accurate beliefs about system reliability: when it tends to be right, when it tends to be wrong, and what factors influence its performance.
Calibration measures become relevant. Can users distinguish high-confidence outputs from low-confidence outputs? Do they adjust their verification behavior appropriately? Do they know when to override the system and when to defer to it?
These are directions for development, not finished frameworks. The specifics will need to be worked out in practice, likely differently for different types of adaptive systems.
Organizational Implications
Expanding the scope of usability work raises questions about organizational boundaries. Traditional usability sits comfortably within design and user research teams. Extended usability work requires closer collaboration with engineering, data science, and product teams who collectively define what the system is supposed to do.
This points to a timing shift as well. For traditional systems, usability evaluation typically happens after the interface is designed. For adaptive systems, usability input may need to begin earlier, when decisions about system behavior are still being made. Helping define the behavioral envelope is different from evaluating the interface that sits in front of it, but both contributions matter. Usability specialists who wait until the interface exists may find that the most consequential decisions have already been made.
What Still Holds
The fundamentals of usability work remain intact. Observation is still essential. Human judgment still matters. Understanding user goals and constraints remains central. Heuristic evaluation, usability testing, and information architecture analysis continue to catch real and important issues.
What changes is the scope of attention. Usability specialists must now consider not only what the system shows, but how it behaves across time and context. The interface remains necessary. For adaptive systems, it is no longer sufficient.
Looking Ahead
As systems become more adaptive, usability work becomes more interpretive. Evaluators are asked to reason about intent, inference, and behavioral consistency in addition to layout and interaction design. This does not require abandoning existing methods. It requires acknowledging their boundaries and extending them thoughtfully.
The future of usability practice lies in recognizing when the interface tells only part of the story, and in developing ways to evaluate the system behavior that users experience but cannot always see.
This shift sets the stage for a closer look at one of our most trusted tools. In adaptive systems, what happens to heuristic evaluation when the system itself begins to think back?