From the lab
Observations, opinions, and dispatches from the work. Essays on AI design, human-centered systems, and what we're learning by building.
4 notes published
Every serious AI deployment ends with the same safety clause: a human reviews the output. That clause assumes the reviewer's expertise is a fixed quantity. Forty years of automation research, and now the first clinical evidence, say it is not.
Most design systems treat accessibility as something you can measure on the visual surface, a contrast ratio here, a labeled component there. The failures that keep the web inaccessible live in the interaction layer those systems never specify.
The DHS Trusted Tester method was designed for people. When you replace the person with an AI, the procedure survives, but the judgment does not transfer. What breaks, and where, turns out to be more instructive than what works.
Most AI products are designed to maximize user confidence, but the research on automation trust says that is the wrong goal. Trust that outruns a system's actual capabilities does not last.