Roguelite Labs

Universal Design for Cognition

1,419w · 7 min · concept, design, accessibility, cognition

Universal Design for Cognition

Ron Mace (NC State, 1980s) developed Universal Design: the design of products and environments to be usable by all people, to the greatest extent possible, without adaptation or specialized design. Seven principles: equitable use, flexibility in use, simple and intuitive use, perceptible information, tolerance for error, low physical effort, size and space for approach and use.

Crucially: these principles were derived from designing for disability but produce better design for everyone. The curb cut effect — see assistive-technology — demonstrates this empirically. Designing for the edges of capability distribution improves the design for the center. Curb cuts were mandated for wheelchair users; they are used constantly by cyclists, delivery workers, parents with strollers, travelers with luggage. The constraint produced a better design for all.

Universal Design for Cognition applies this logic to cognitive systems: design AI and cognitive tools to be usable by people with the widest possible range of cognitive architectures — different literacy levels, different linguistic backgrounds (future-time-reference), different working memory capacities, different spatial reasoning profiles — and the resulting design will be better for everyone.

This is not accessibility as afterthought but accessibility as design driver. The thesis claim: designing AI for people with low literacy or non-dominant FTR linguistic backgrounds will produce interfaces that make better cognitive scaffolds for everyone, for the same reason that curb cuts produce better surfaces for everyone.

Six Derived Principles

1. Scaffold, Don't Substitute

From extended-cognition and assistive-technology: tools should build internal cognitive capacity over time, not permanently externalize cognitive operations. Verify through use that users are internalizing scaffolded operations, not becoming dependent. Design periodic reduction of scaffolding to test internalization — Vygotsky's zone of proximal development as a design principle rather than a developmental observation.

A reading tool that reads to a user is substituting. A reading tool that reads to a user while teaching phonemic awareness is scaffolding. The cognitive output is different even when the immediate product is the same. Most AI interfaces are not designed to make this distinction; they are optimized for task completion, which is the right metric for automation and the wrong metric for augmentation.

2. Calibrated Coupling Support

From extended-cognition and hci-ai: design for appropriate trust — neither over-trust nor under-trust. Expose system reasoning when relevant. Signal uncertainty explicitly. Make verification pathways visible. Preserve human situation awareness even when the AI acts autonomously.

A user who cannot calibrate their coupling cannot extend their cognition effectively. Over-trust produces automation bias: users ratify AI outputs without independent evaluation. Under-trust produces friction loops that eliminate the productivity benefit. Calibration requires information — about confidence, about scope, about what the system did not know when it generated the output. Most current AI interfaces provide almost none of this.

The universal design angle: calibration support is most urgently needed by users with the least prior AI experience and the least domain expertise. But the same calibration infrastructure — explicit uncertainty signals, surfaced reasoning, bounded scope disclosure — improves the interface for expert users too. They just suffer less obviously without it.

3. FTR Prosthetics

From future-time-reference: externalize temporal structure for users whose linguistic background hasn't built strong future-state representational schemas. Strong-FTR languages grammatically obligate explicit future-state encoding on every future-referring utterance; weak-FTR users have had systematically less practice with this cognitive operation through normal language use.

AI interfaces can provide temporal scaffolding structurally: step-by-step decomposition prompts, explicit goal-state specification fields, timeline scaffolding, before/after state descriptions. These structures do not require the interface to know the user's FTR background — they are universally available scaffolds that users with well-developed future-state schema can ignore and users who need them can use.

The universal benefit: explicit temporal structure improves task specification quality for all users. The users most obviously helped are those with weaker default future-state schemas; the users also helped are everyone else planning complex tasks who benefits from being prompted to articulate the end state before specifying the path.

4. Literacy-Independent Access with Literacy-Building Design

The central goal of assistive-technology applied to AI: reduce the immediate literacy prerequisite for accessing cognitive extension while actively building literacy over time. This is the difference between a read-aloud tool that also teaches phonics and one that only substitutes for reading.

AI interfaces can be designed to model good specification, show users what well-formed prompts look like, and gradually transfer more of the specification work to the user. The interface doesn't require initial literacy to use, but it is structured to produce literacy growth over time. The scaffold is designed to dissolve — not permanently compensate.

This reframes the current debate about AI and literacy. The concern is that AI substitutes for literacy. The design response is not to restrict access — it is to design AI interfaces that build literacy as a side effect of use. The question is not whether people will use AI; they will. The question is whether the interface is designed to make them more or less capable over time.

5. Situated Awareness Preservation

From Suchman's situated action framework (hci-ai): users are situated agents, not plan-followers. Plans are resources for action, not programs that execute. Users encounter unexpected features of their situation mid-task and update their intentions accordingly. AI systems that model users as operating from fixed goals, and that fail or produce degraded output when the situation changes, produce frustrated users who override the AI at exactly the moments when AI assistance would be most valuable.

Preserve user ability to update, interrupt, and redirect the AI's activity in response to changing situational needs. Design AI that expects situational updating rather than treating it as user error.

This is a universal design principle because situatedness is not a special-case user property — it is universal human cognition. All users are situated agents. AI interfaces that don't accommodate situational updating are failing all users; the failure is most visible with users whose situations are most different from the interface designer's implicit model.

6. Transparent Capability Space

From Norman's discoverability problem (hci-ai): if users cannot perceive what's possible, the most capable users are the most advantaged — because they already know the capability space through prior experience, adjacent expertise, or social networks that share that knowledge. Users who don't already know what to ask for are systematically underserved.

Designing for users who don't know what to ask for — low familiarity users, users from different linguistic and cultural contexts, first-generation technology adopters — requires explicit capability surfacing, example-rich interfaces, and structured entry points that don't require users to already have expert mental models.

Universal benefit: most users at most times, even expert users, don't know what's optimally possible in a new system. Good discovery design helps everyone. The users most obviously helped are the ones with no prior model; the ones also helped are experts navigating new capability territory.

The Policy Dimension

If AI interfaces function like FTR grammar — structurally obligating certain cognitive operations through their design — they are a cognitive policy instrument. Interface designers are making decisions about what billions of people practice thinking, at scale, without recognizing the design as cognitive policy.

When a grammatical structure encodes future-state thinking on every utterance, it trains a cognitive pattern across a language community over generations. When an AI interface structures interaction around goal-state specification, it trains (or fails to train) a cognitive pattern across its user base in real time. The scale difference is large; the mechanism is structurally similar.

Universal Design for Cognition is the framework for making those decisions consciously. The highest-stakes version: if the AI interface externalizes the cognitive operations that determine economic and collaborative behavior — per Chen's FTR finding that linguistic future-time reference predicts savings rates and health behaviors — then interface design choices propagate cognitive patterns at the scale of language itself. This is not hyperbole. It is the logical extension of the future-time-reference finding to the current technological environment.

The design implication is not that every interface must optimize for every user. It is that cognitive equity should be an explicit design objective, not an afterthought, and that the tools for achieving it — the six principles above — are available and well-grounded in existing theory.

Related

extended-cognition · assistive-technology · hci-ai · future-time-reference · tools-for-thought · personalized-systems · lot-llm-paradox

Sources

linked from
Assistive Technology