Google Health Coach Invented a Workout. Here's Why AI Fitness Apps Do This.

Large language models do not know what you did yesterday. They predict the next token based on the patterns in their training data. When you ask a fitness coach chatbot "what have I been up to this week?", the model needs your workout data injected into the context window before it can answer accurately.

If that injection is missing, incomplete, or noisy, the model does not say "I don't know." It generates a plausible-sounding response. Active users run. Active users run about 3-7 miles. So the model produces a run that fits the pattern. This is not malice. It is how next-token prediction works when the grounding is weak.

All three conditions were likely present in the Google Health Coach incident. The coach correctly referenced real data in some places (sleep, a real workout), which means grounding was partial. The 5-mile run appears to have been the model filling a gap it could not verify.

Google Health is an aggregator. It pulls from your Fitbit, Apple Health, MyFitnessPal, Peloton, connected apps, manual logs, and health records. This breadth is one of its genuine strengths for biometric tracking. But it creates a specific challenge for AI summarization.

When 6 data sources feed into a context window, each with different schemas, update frequencies, and reliability levels, the resulting context is messy. Gaps appear between sources. A step count from the wearable might not align with a workout from a connected app. An inferred calorie burn from the phone accelerometer exists alongside a manual food log entry.

The model has to reason across all of this. When it cannot reconcile a gap, it fills it. A specialized app that only tracks what a user explicitly logs has a narrower but cleaner context. The model has less to hallucinate across.

Grounding means attaching verified data to the model's context before it generates a response. In fitness, that means: before the AI answers a question about your training, it reads your actual logged sets, reps, weights, exercises, and dates.

Research on retrieval-augmented generation (RAG) systems consistently shows grounding reduces hallucination rates significantly. A 2026 developer guide on LLM grounding (neuledge.com, published February 2026) cites 42-68% reduction in hallucinations when models are given retrieved real context vs no context. Clinical RAG systems have pushed hallucination rates to 5.8% using self-reflective verification layers on top of retrieval.

Zealova's workout plan generator and coach do this: before any AI call related to your training, the system pulls your actual logged workout history and injects it as structured context. The model is not asked to guess what you have been doing. It reads what you logged.