HCI in the Age of AI

Classical HCI developed around the WIMP paradigm — Windows, Icons, Menus, Pointing — and the theoretical apparatus built to explain it was calibrated to that paradigm's assumptions. AI-mediated interfaces break nearly every one of those assumptions: there is no stable visual object to manipulate, the system's capabilities are not enumerable by looking at it, and feedback is delayed and probabilistic rather than immediate and deterministic. The field's foundational concepts need systematic re-examination, not incremental extension.

The WIMP Paradigm and Its Assumptions

The WIMP paradigm, formalized in the 1980s through work at Xerox PARC and codified by Ben Shneiderman's direct manipulation principles (1983), rests on three structural assumptions: (1) the system's state is visually represented at all times; (2) physical actions correspond directly to operations on visible objects; and (3) feedback is immediate, legible, and reversible. These assumptions enabled a generation of interaction theory — affordances, mental models, the action-evaluation cycle — that was internally coherent and highly productive.

Conversational AI violates all three. There is no persistent visual representation of system state; the system's capabilities are implicit and unbounded; feedback is neither immediate nor unambiguous. Interface theory built on WIMP assumptions does not straightforwardly apply, and extending it by analogy risks preserving the wrong abstractions.

Norman's Gulfs — What AI Changes

Don Norman's The Design of Everyday Things (1988) defined two cognitive gaps that good interface design should minimize:

• Gulf of Execution: the difficulty of determining how to make the system do what you intend
• Gulf of Evaluation: the difficulty of determining what the system did and whether it matched your goal

In WIMP interfaces, the Gulf of Execution is narrowed by affordances — visual and tactile cues that signal available actions. A button signals pressability; a slider signals range. The user need not know what the system can do in general; they need only perceive the options currently available. In conversational AI, natural language theoretically collapses this gulf — you describe what you want in ordinary language. But this apparent improvement conceals a new execution problem: discoverability. Users have no perceptual cues about what the system can do. What was a narrow, bounded problem (how do I operate this specific button?) becomes a wide, open problem (what is even possible to ask for?).

The Gulf of Evaluation deepens more severely. AI outputs are typically fluent and confident regardless of their accuracy. The low-level feedback signals that humans rely on for calibration — error messages, disabled controls, visual state changes — are absent. The interface provides no signal distinguishing correct from plausible-sounding-but-wrong. A hallucinated citation looks identical to a correct one; a subtly wrong answer is indistinguishable in surface form from a correct one.

Suchman and Situated Action

Lucy Suchman's Plans and Situated Actions (1987) is the most important HCI text for the AI age. Her target was the planning paradigm in AI systems — the assumption that human action can be modeled as the execution of pre-formed, hierarchically organized plans. Her argument: human action is fundamentally situated, improvised in response to contingent and emergent circumstances, not the execution of plans. Interactive systems that model users as plan-followers fail because users are not plan-followers; they are responsive, situated agents whose actions are shaped by the concrete details of the situation as it unfolds.

Her specific account of AI assistant failures is precise: they fail when the user's actual situation diverges from the situation the model has implicitly constructed. The system always operates from an implicit model of the user's context — and that model is always incomplete. The gap between the model and the actual situation produces the characteristic failures of AI systems: technically correct responses that are practically wrong because they address the wrong situation. Suchman's framework predicts this systematically; it is not a quality problem, it is a situatedness problem, and it cannot be solved purely by improving model capability.

Situation Awareness (Endsley)

Mica Endsley's three-level model of situation awareness, developed from aviation and military systems research, provides the right vocabulary for understanding AI supervision failures:

1. Perception: detecting the elements present in the environment
2. Comprehension: understanding what those elements mean for current goals
3. Projection: anticipating future states and events

Endsley's core finding: automation failures are not random. They follow from the degradation of situation awareness that automation itself causes. When a system appears to be operating correctly, humans stop monitoring it — they lose perception. When it fails, they lack the comprehension and projection needed to diagnose and recover. The automation that removes the cognitive work also removes the cognitive engagement that makes supervision possible.

The direct implication for AI systems is uncomfortable: AI tools that take cognitive tasks away from humans may degrade the situation awareness needed to supervise those tools. extended-cognition requires calibrated coupling between the human and the external system; good situation awareness is what calibrated coupling looks like in practice. Automation that produces SA degradation produces a system that can fail catastrophically and without warning.

Recognition-Primed Decision Making (Klein)

Gary Klein's naturalistic decision making research, documented in Sources of Power (1998), showed that experts under real conditions do not deliberate analytically between options. They pattern-match the current situation to a recognizable situation type, generate a single candidate action based on that recognition, and mentally simulate whether it will work. Only if the simulation fails do they generate an alternative. This is the Recognition-Primed Decision (RPD) model.

The design implication for AI is counterintuitive: presenting users with multiple options to compare imposes a deliberative cognitive mode that expert users do not normally employ, and which is slower and less accurate than recognition-based evaluation. AI interfaces that generate ranked alternatives, comparison matrices, or multiple candidates optimize for a cognitive strategy that real experts rarely use. Better to present a single strong candidate — the model's best attempt — and let the expert evaluate it against their situational understanding. This matches expert cognitive style; it preserves the RPD workflow rather than replacing it.

The Discoverability Problem

Gibson's ecological affordances — physical properties of objects that signal available actions — were adopted by Norman as the core mechanism for reducing the Gulf of Execution. The affordance is perceptual: you see that the cup is graspable, the door is pushable, the button is pressable. Affordances work because action possibilities are visible in the object.

Conversational AI has no affordances. You cannot perceive what a language model can do by looking at it or at a prompt box. The entire capability space must be discovered through use, external documentation, or word of mouth. This creates a structural inequality in AI tools: advanced users who have built a model of the capability space through accumulated experience get dramatically more value than novice users — not because the interface has been designed for experts, but because expertise is the only available path to navigating an undiscoverable space.

This is the field's most urgent unsolved design problem. AI interfaces need new mechanisms that serve the function affordances serve in physical interfaces: communicating action possibilities without requiring the user to already know what to ask for. Example prompts, structured templates, and dynamic capability surfacing are partial approaches, but they do not yet constitute a coherent affordance theory for conversational capability spaces.

The Verification UX Problem

extended-cognition argues that productive human-AI coupling requires calibration — the human must maintain appropriate trust in AI outputs, neither over-trusting nor under-trusting. Current AI interfaces provide almost no support for calibration. Outputs are presented uniformly regardless of the system's internal confidence. The reasoning behind outputs is typically hidden. There is no mechanism for the system to signal uncertainty in ways that users can act on.

The result is predictable: users calibrate trust through heuristics — prose quality, plausibility, surface fluency — that are orthogonal to accuracy. A confident hallucination and a confident correct answer are visually identical. A hesitant correct answer may be dismissed. The interface is systematically anti-calibrating.

Designing for calibrated coupling means: expose reasoning when it is decision-relevant, signal uncertainty visually and linguistically, make verification pathways explicit, and preserve human situation awareness even when the AI is operating autonomously. This is not primarily a model problem. It is an interface design problem, and the tools to address it are largely available. What is missing is a design culture that treats calibration as a first-class requirement.

Key Figures

• Don Norman — The Design of Everyday Things (1988); affordances, mental models, the Gulfs framework; Human-Centered AI (2022) returns these concepts to the AI context
• Lucy Suchman — Plans and Situated Actions (1987); the foundational critique of planning-paradigm AI and its implications for interface design
• Mica Endsley — situation awareness research; Designing for Situation Awareness (2003); the SA degradation model for automation failures
• Gary Klein — naturalistic decision making; Sources of Power (1998); the RPD model and its implications for expert-supporting interface design
• Edwin Hutchins — distributed-cognition; system-level cognitive design; Cognition in the Wild (1995)
• Ben Shneiderman — direct manipulation (1983); now a primary voice on human-centered AI; Human-Centered AI (2022)

Related

extended-cognition · distributed-cognition · tools-for-thought · agentic-workflows · computer-use

Sources

• Norman, D. (1988) The Design of Everyday Things, Basic Books
• Suchman, L. (1987) Plans and Situated Actions, Cambridge University Press
• Endsley, M. (2003) Designing for Situation Awareness, CRC Press
• Klein, G. (1998) Sources of Power: How People Make Decisions, MIT Press
• Shneiderman, B. (2022) Human-Centered AI, Oxford University Press
• https://www.jnd.org/