Google researchers outlined an experimental way to turn Fitbit histories into 12-week AI coaching plans on November 11, 2026. For wearable users considering personalised wellness advice, the key distinction is what was evaluated: experts assessed the plans, but the study has not shown that following them improves health.
The coaching preprint was submitted on November 5. Google describes the models as research prototypes, not diagnostic tools. Its announcement gives no consumer release date or pricing for this particular system.
From wrist measurements to a what-if plan
The framework connects three stages: wearable measurements, simulated futures and written advice.
Google’s explanation starts with Coachable Health Age: a score combining resting heart rate, steps, sleep duration and Active Zone Minutes, an activity-intensity measure. It expresses selected health-risk signals in age-like units. Google reports that the score improved disease-risk prediction in observational follow-up. That does not show what happens when someone acts on a coaching plan.
A forecasting model called SensorWM uses two baseline weeks to simulate the next 12 weeks under different habits. A language model then turns selected scenarios into a coaching plan. “Counterfactual” means a modelled what-if alternative—not an observed future.
The accompanying heart, kidney and metabolic-health study is separate: it combined wearable measurements with blood biomarkers in 819 people. It does not establish that a wrist sensor alone can diagnose kidney disease.
What the expert ratings mean
In the paper’s offline evaluation, five coaches and five clinicians assessed 49 participant profiles, producing 245 paired comparisons per panel. Coaches preferred simulation-informed plans in 175 comparisons (71.4%). They favoured them for expected adherence in 174 (71.0%): a judgment about likely follow-through, not measured adherence.
Both versions used the same language model; one received extra health-age and simulation information. This was not AI competing against a human coach.
Clinicians endorsed exercise-intensity limits in 243 of 245 ratings (99.2%), versus 239 (97.6%) for comparison plans. The difference was not statistically significant. Neither percentage measures harm during actual use.
The authors’ central limitation: SensorWM learned observational associations, which can be affected by unmeasured factors such as medication changes or illness. They say randomised trials are needed to establish real-world benefits.
A lower health age is not extra lifespan
The practical implication of an age-labelled risk score is easy to miss. Suppose a forecast showed a two-year reduction—an illustrative example, not a result reported here. That would be a projected change in the score, not evidence of two additional years of life or literal reversal of ageing.
The FDA’s explanation of clinical endpoints helps clarify the distinction: a marker can stand in for a health outcome only with evidence supporting that use. Showing that people experience better health or longer survival is different from improving a substitute measure. This FDA resource concerns drug and biologic development; it is not an assessment of Google’s score.
For the broader scientific distinction, see why a good fit to data is not enough to explain what will happen.
Three checks before relying on a forecast
For someone considering an AI wellness service—or evaluating one for employees—the practical evidence check is to keep three questions separate:
What was measured? Separate recorded sensor readings from simulated changes and expert opinions about advice.
Did people follow the plan? A professional’s expectation that advice is achievable cannot establish actual follow-through.
What improved? Look for measured outcomes in people using the specific system, not just an attractive score or a large training dataset.
For this study, the practical conclusion is narrow: experts often preferred the simulation-informed plans. Whether users can sustain them—and benefit from them—remains a question for real-world trials.