Gemini in Google Home Keeps Mistaking My Dog for a Cat (www.wired.com)

Google rolled Gemini into the Home app to bring large-language-model smarts to Nest cameras and smart-home controls—replacing Assistant on many devices and adding Ask Home, Home Brief recaps and richer camera alerts (e.g., “FedEx dropped off two packages” instead of “Person seen”). The author liked package detection and natural-language automations, but after enabling Gemini for Home it repeatedly misclassified their dog as a cat in live alerts and daily summaries. Telling Ask Home “that’s a dog, not a cat” was acknowledged but didn’t stop the errors; Google says these features are early-access, that it’s investing in more accurate identification (including using user-provided corrections), and that improvements hinge on the underlying Familiar Faces system, which today doesn’t support pets. For the AI/ML community this is a useful real-world stress test of multimodal LLM deployments: improved natural language and automation can mask persistent vision-model failures, label hallucinations, and imperfect user-feedback loops. Key technical takeaways are the need for better pet/object class coverage, robust fine-tuning or continual learning pipelines that actually incorporate user corrections, and evaluation on edge/cloud camera variability. There are also practical implications around false alarms, user trust, and privacy when scaling neighborhood-level features (e.g., Ring’s Search Party). The episode highlights that productionizing multimodal AI still requires careful dataset, feedback-loop, and UX design to avoid obvious but sticky misclassifications.