An informal, human-centered heuristic for judging an AI system's elegance, reliability, and user delight by focusing on simplicity, predictable behavior under stress, and ease of recovery when errors occur.
The Wozniak test is a practitioner-oriented evaluation concept that prioritizes user-facing qualities of AI systems (notably clarity, minimal surprise, and robust recoverability) over purely capability-centric metrics. Unlike the Turing test, which probes anthropomorphism, the Wozniak test asks whether an AI's design and behavior feel thoughtfully engineered from a user's perspective: are its failure modes understandable, is it easy for users to recover from mistakes, does the system expose the minimum necessary complexity, and does it avoid needless edge-case brittleness that damages trust? In engineering terms this translates into measurable properties such as calibration and uncertainty communication, failure-mode taxonomy and mitigation, latency and feedback loop characteristics, interpretability affordances, and UX metrics (task success, error recovery time, subjective trust). For AI/ML (Machine Learning) systems this heuristic guides model selection, interface design, and acceptance criteria—favoring simpler, better-calibrated models or modular hybrid architectures when those produce more predictable, debuggable behavior for end users, and shaping evaluation suites that combine behavioral tests, adversarial and distribution-shift trials, and human-in-the-loop recovery scenarios.
No formal origin is documented; the phrase has appeared informally in tech and AI/UX discussions since the late 2010s and gained traction among AI product, UX, and safety practitioners around 2020–2024 as concerns about trustworthiness, interpretability, and human-centered design intensified.
There is no single inventor of the concept; it is named by analogy to Steve Wozniak's design ethos (simplicity and user delight) and has been shaped by contributions from the HCI community (e.g., human-centered design advocates such as Don Norman), AI safety and interpretability researchers and teams (academic labs and industry groups such as those at major AI research organizations), and AI product/UX practitioners who operationalize reliability-and-recoverability checklists for deployed systems.