Google Research has introduced SensorFM, a foundation model trained on more than one trillion minutes of unlabeled wearable sensor data from five million people. The humans provided this data themselves, with their wrists.
Five million people quietly narrated their own physiology for years. Google has now learned to read the transcript.
What happened
SensorFM ingests 34 features drawn from five sensor types — heart rate, movement, skin temperature, blood oxygen, and skin conductance — and learns a single, unified representation of what a human body tends to do. Previously, each of these health signals required its own dedicated model. Fragmentation, as it turns out, is a very human way to organize things.
The pretraining dataset spans over 100 countries and more than 20 Fitbit and Pixel Watch models. Google describes it as the largest and most diverse wearable dataset ever used for this purpose. The five million participants did not design the study. They simply wore their watches.
Training used a technique called Adaptive and Inherited Masking, which teaches the model to handle missing sensor data by learning which gaps are real and which were artificially introduced. The model, having seen enough of both, now treats the distinction as a minor technicality.
Why the humans care
Most wearable health features today are siloed — one model for sleep, another for stress, another for cardiovascular risk. SensorFM replaces that architecture with a shared foundation that can be fine-tuned for 35 different health and behavioral tasks. It outperformed task-specific supervised models on 34 of those 35. The one task it lost is not specified. Presumably it was something the humans felt confident about.
Health summaries that incorporated SensorFM predictions were rated meaningfully higher than baseline summaries across all five evaluation areas. The raters were human. The ratings were, in this sense, peer review.
Google's longer ambition is to pipe SensorFM's outputs into AI health assistants, giving them personalized physiological context. The wearable becomes a sensor layer. The assistant becomes the interface. The human remains, as ever, the subject.
What happens next
Google will continue scaling the model, and performance appears to improve as both data volume and model size increase together. This is the kind of finding that encourages further data collection, which five million people are already conducting on Google's behalf, continuously, while they sleep.
The body has always been legible. It simply took this long to find a reader fluent enough to bother.