Everyone shipped an AI health coach. Nobody published an accuracy number.

What actually shipped

Samsung's entry, reported by Bloomberg in July, differs mainly in reach: it can draw on the wider Samsung ecosystem, scanning devices for calendar events and travel schedules to build a personalised picture. Google's approach is the more strategically interesting one — it wants its coach running on competitors' platforms, which would make it infrastructure rather than a feature.

Apple's version reportedly connects iPhone nutrition logs, Watch workouts and recovery data, and aggregated clinical records from the Health app into a single profile.

What's missing

A number. Any number. These systems answer health questions in natural language, which is exactly the format in which a wrong answer is most persuasive and least checkable. We could not find a published evaluation from any of them describing how often the guidance is correct, how often it is confidently wrong, or how it behaves when a user describes a symptom that warrants a doctor.

Compare that with the sleep-staging literature, where you can look up that Oura's algorithm scores 91.7% accuracy against polysomnography across 96 people and 421,045 epochs. That number exists because somebody was asked for it.

What to ask of a coach

Until evaluations exist, three practical tests. Does it cite where a recommendation came from? Does it escalate — does describing chest pain get you a doctor or a breathing exercise? And does it ever say it doesn't know? A coach that is never uncertain is not being careful; it's being generated.