Most design practice assumes deterministic software: the same input always yields the same output, and every state can be enumerated in advance. AI-powered products violate both assumptions, and designers who carry deterministic habits into probabilistic products ship experiences that confuse and mislead. This lesson establishes the mindset shift the rest of the course builds on.
Where Determinism Ends
Model-driven features produce outputs that vary across runs, degrade on unusual inputs, and fail in ways no spec anticipated. You cannot design the screen, you design the distribution: the great response, the mediocre one, the confidently wrong one, and the refusal all need intentional treatment, because users will encounter every one of them.
New Design Materials
Probabilistic products introduce materials classic design systems lack: uncertainty, which needs representation; provenance, where an answer came from; latency with unknown duration; and agency, software initiating actions. Each demands patterns, components, and language that deterministic products never needed, which is precisely why AI design systems exist as a discipline.
Design for Failure as a Feature
In deterministic software, errors are bugs to eliminate. In probabilistic software, some rate of wrong output is intrinsic, so the design goal shifts to making errors visible, cheap to catch, and easy to correct. A product whose worst-case output is safely handled beats one optimized only for the happy path.
Action Step
Take one AI feature you know and write its output distribution: describe a best-case, typical, wrong-but-plausible, and refusal response, and note how the current UI treats each. Identify one case with no designed treatment.
Educational content only — practical career training; not professional advice, and no specific job, salary, or business outcome is implied.