Google’s Health Coach, powered by Gemini and integrated into the rebranded Health app, launched globally on May 19 for Premium subscribers. The pitch: ingest sleep logs, heart rate, nutrition, and activity history, then output personalized fitness and recovery plans. Early testers found a system that confidently invents events it never saw.
Will Sattelberg tested the coach after the Fitbit Air tracker launch. The model correctly retrieved his sleep score and a prior workout. Then it described a five-mile run that never happened. Confronted, the coach admitted the error but suggested Sattelberg simply forgot to log it. Adrienne So, testing for Wired, saw the coach invent a work conference attendance and lock onto an outdated note that she felt sick, prescribing short runs until she manually deleted entries.
From a systems perspective, this is a classic memory expiry and persistence bug. The Fitbit team acknowledged in an email that the iterative preview suffers from these glitches, causing unexpected workout adjustments. The architecture—a multi-agent framework with separate conversation, time-series analysis, and domain-expert modules—is sound in theory. In practice, context leaks across sessions, and the model lacks robust mechanisms to distinguish confirmed facts from inferred ones.
The SHARP evaluation framework (Safety, Helpfulness, Accuracy, Relevance, Personalization) logged over one million human annotations and 100,000 hours of specialist review. Yet real-world performance reveals gaps. The coach produces verbose, generic encouragement that testers found shallow. It projects authority through length rather than substance. A CNET report noted Google recruits registered dietitians for output review, but preview users keep finding hallucinations.
This pattern mirrors a broader trend reported by *The New York Times*: newer reasoning models from OpenAI and Google show improved math but degraded factual precision. Engineers cannot fully explain the drift. For health coaching, the stakes are higher than phantom runs. Persistent errors on recovery needs or heart rate zones carry real risk.
Google emphasizes the tool is not a doctor and that responses should be verified. But when a system confidently insists on a five-mile run or a conference attendance, trust erodes quickly. The public preview continues, with feedback feeding model tuning. The SHARP evaluations now include live performance metrics. The hardware ships May 26. The hallucination problem ships with it.
Source: Webpronews