In 2025, the promise of generative AI transforming the smart home has largely fallen short of expectations. Despite years of hype, many users—including myself—are finding that AI-powered assistants like Alexa Plus and Google’s Gemini for Home are less reliable than their predecessors when it comes to basic tasks. I asked my Alexa-enabled Bosch coffee machine to make me coffee this morning, and it refused—again. Not because the machine was broken, but because the AI assistant couldn’t execute the simple command, offering a different excuse each time. This isn’t an isolated issue. The shift from rule-based, command-driven assistants to conversational, LLM-powered systems has introduced a new layer of unpredictability. While the new assistants are more natural in conversation, better at understanding complex requests, and capable of handling tasks like managing calendars or suggesting recipes, they struggle with fundamental smart home functions—turning on lights, running routines, or controlling appliances. The root of the problem lies in the fundamental difference between old and new AI architectures. Earlier voice assistants worked like template matchers—recognizing keywords and triggering predefined actions. They were predictable and reliable, even if limited. Generative AI, however, operates with stochasticity, or randomness. The same command can produce different responses, and the assistant may overthink simple requests, leading to errors. Experts like Mark Riedl from Georgia Tech explain that LLMs aren’t designed for the repetitive, deterministic tasks that smart home systems require. Instead of waiting for a keyword, they must now generate correct API calls on the fly—requiring precise syntax and context. This adds complexity and increases the chance of failure. “It’s not just about understanding language,” Riedl says. “It’s about generating the right code to make the device respond.” Companies like Amazon and Google have tried to bridge the gap by layering multiple models—using a more constrained system for reliability and a more powerful one for conversation. But this hybrid approach often leads to inconsistency. Users may get a smooth response one time and a failure the next, even with the same command. The tradeoff is clear: greater flexibility and intelligence come at the cost of reliability. As Dhruv Jain from the University of Michigan notes, companies are prioritizing innovation over stability. “They release fast, collect data, and improve over time,” he says. That means consumers are effectively beta testers for AI systems that aren’t ready for prime time. While the long-term vision remains compelling—a truly proactive, intelligent home that anticipates needs and chains tasks together—today’s reality is frustrating. If AI can’t reliably turn on the lights, how can we trust it with more complex responsibilities? The path forward likely involves refining how LLMs balance precision and creativity. But until then, the smart home isn’t smarter—it’s just more unpredictable. And for many of us, that’s a step backward.