A 1980s Xerox photocopier study revealed crucial lessons for today's AI user experience. Researchers observed users struggling with a copier's rigid instructions when faced with unexpected jams. This highlighted that human behavior is improvisational and context-dependent, not a predictable script. Early HCI development at Xerox PARC also saw a tension between system-centric and human-centric design philosophies. While some believed smarter machines were inherently more useful, studies demonstrated the importance of accounting for real-world complexities. Anthropologist Lucy Suchman’s work showed that human action is a constant adjustment to circumstances, not a fixed plan. Similarly, research on collaborative systems found that rigid communication models fail because human interaction is a dynamic dance. These early insights led to principles like situated action and person-centered design. Traditional systems struggled to meet these requirements because they lacked context and flexibility. However, modern Large Language Models (LLMs) are beginning to address these, offering probabilistic reasoning and interactive error correction. This shift is enabling a new interaction paradigm called AIX, where users state intent and the system proactively reasons. Despite their fluency, LLMs create an illusion of human competence by masking their limited understanding of physical reality. This leads to "jagged capabilities," where AI excels in some areas but fails unexpectedly in others. Current conversational interfaces often exacerbate trust issues by lacking visible affordances and ignoring multi-stakeholder needs. Building reliable AI systems requires understanding trust as an ongoing negotiation within a network of users and their diverse needs.