Google’s Health Coach Keeps Inventing Workouts You Never Did

When Google launched its AI-powered Health Coach this month, the pitch was simple: analyze your sleep, heart rate, nutrition, and workout history, then deliver personalized fitness plans. Early testers got something else—a system that confidently fabricates exercise sessions.

Will Sattelberg, testing the feature with the new Fitbit Air tracker, watched the AI correctly pull his sleep score and a real workout. Then it described a five-mile run he never took. When challenged, the coach apologized but suggested Sattelberg simply forgot to log it. Adrienne So, writing for Wired, found the coach insisted she was attending a work conference she wasn’t at. Later, it locked onto an old note that she felt sick and kept prescribing short, slow runs—even after she corrected the record. She had to manually delete entries to reset the behavior.

Fitbit’s team acknowledged the memory glitches in an email, calling them expected in an iterative preview. They said fixes are underway. But the pattern raises questions about readiness. Google rolled out the coach globally on May 19 for Health Premium subscribers, pulling data from Fitbit devices, lab uploads, and manual logs. The company stresses the tool is not a doctor, and responses should be checked for accuracy.

These reports arrive as major LLMs from OpenAI and Google show more confident mistakes on specific facts. Health coaching demands precision. A phantom run might seem minor, but persistent errors on recovery needs or heart rate zones carry higher stakes. Google has invested heavily in safeguards—a multi-agent architecture, a Consumer Health Advisory Panel, and the SHARP framework backed by over 100,000 hours of specialist review. Still, preview users keep finding hallucinations.

Competitors like WHOOP are leaning into human clinicians instead. Google promises ongoing iteration, and feedback flows directly into model tuning. But trust erodes fast when an AI invents workouts. The coach may help some users run consistently. For others, it’s a reminder that even careful engineering can’t yet silence the ghosts.