I’m Dr. Dave Gilbert. For thirty years I’ve studied one thing under many names: how people make sense of, break down with, and repair their understanding of intelligent machines — and what design owes them when the plan runs out.
It is a single lens, held steady from a 1995 conversation-analytic usability study to instrumented research on today’s AI systems. The speculative essays on this site are not a separate project from that research; they are its latest application. Repair, coordination, and the interface as a governed contract are not metaphors I reached for — they are findings I have been living inside since the beginning.
Usability is a function of the resources available to users for the repair of breakdown. Breakdown is inevitable; good design anticipates it and provides for its repair.
— from the 1995 study, and still the operating principle.
The method draws on conversation analysis, Lucy Suchman’s situated action, Edwin Hutchins’s distributed cognition, Goffman’s frames, Winograd and Flores on breakdown, and Goodwin & Goodwin on gesture — traditions I committed to in 1995, have extended ever since, and that have never been more explanatory than they are now. Read the method →
Study the encounter, not a decontextualized score. What something means is displayed in the next turn.
Friction, repair, and frame-loss reveal the system. A breakdown unconceals the web of tools and expectations that worked silently until it didn’t.
The gap between what people say and what they do. Polite words mask friction; the job is to catch the delta.
Instruments shape what they measure and categories presuppose their phenomena. Make the seams visible instead of smoothing them away.
A clean, consciously organized ontology is the ground of usability — the background of obviousness every action presupposes.
The lens didn’t stop at studying AI; lately I build with it — and two of the systems are now live in public beta: Suminar and Mem·Sum, with Mail·Sum under construction (the case study). The systems I make — a relationship-memory layer, a personal knowledge base, a scholarly-retrieval server — all follow one architecture rule, and it is the research talking. Keep the tool (an MCP server, in current terms) a thin, deterministic kernel that protects what software is good at — immutable records, stable IDs, provenance, atomic writes, permission scope — and lean on the semantic capacity the user already brings: BYOC, bring your own cognition. No type enums, no get_inbox, no precomputed views; the model reads the markdown and computes the view. The ontology isn’t declared up front in a schema — it emerges from use, the same conversation-analytic move that runs the rest of this practice. It is the builder’s form of “meaning is displayed in the next turn.” The same provenance-first discipline has also shipped publicly: CiteMap, the AI-powered archive behind the FBT Voices project (see About), keeps every summary pointing back to its original source.
Do not make the tool semantic. The chatbot already is.
— the rule under everything I build. Read the essay →
One method, hundreds of projects. These seven case studies trace its arc across escalating substrates — listed newest first, from today’s shipped products back through the AI interface, the usability lab, the aging home, the retail aisle, and the museum to a pocket PDA in 1995.
The argument turned into live products, run as research probes: Suminar’s source agents and Mem·Sum’s shared Sums in public beta, Mail·Sum under construction. No house chatbot, kernels that don’t think, operators content-blind by construction — and the beta learns only what people choose to share.
The same lens turned on AI-mediated experiences: where polite speech hides friction, where an AI’s claim of action outruns its execution, and where commerce wants to be a forum rather than a monologue. Described here by method and role only.
A multi-track lab that records a usability session as parallel, time-aligned tracks — speech, screen, and a physical feedback controller treated as a non-verbal turn rather than a measurement. PRISM is Patrick Squire’s concept and SentiMeter hardware; I developed the philosophy and methodology, and we hope to write it up together.
An aging-in-place system that read everyday appliance use as a sign of well-being. Its hard problem was representational: sensors shape phenomena, and inferring “grandmother is okay” from a coffee pot demands seamful design — visible seams, not a false-smooth surface.
A mock retail environment — slatwall, fixtures, packaging, adjacency — let shoppers decide in the situation that actually produces the decision. Distributed cognition pointed at Walmart, Apple Retail, Best Buy, and Target.
With my students, unofficial iPod audio guides and an SMS gallery-comment system layered an alternate, contestable voice over MoMA — a user-carried device rewriting an institution’s interpretation. Covered in the New York Times.
Two novices set up email on an early agent-based PDA. Read turn by turn, the session shows that consistency of frame is paramount and that usability is a function of the resources available for the repair of breakdown — the sentence that still runs the whole practice.