Google says its trillion-minute wearable AI matches lab tests. No one else has checked.

Google Research says a foundation model it trained on roughly a trillion minutes of consumer smartwatch data can predict 35 health conditions as accurately as a clinical blood panel. Every public number in the SensorFM paper is the company's own. No independent team has run a comparable benchmark.

The model went up on arXiv in May 2026 and was accompanied by a Google Research Blog post from Senior Research Scientist Xin Liu and Staff Research Scientist Daniel McDuff. Lead author Girish Narayanswamy is listed with 39 co-authors from Google Research, Google DeepMind, and academic institutions. The 35 endpoints include cardiovascular risk, depression, and anxiety, alongside 32 others the paper does not enumerate in the abstract.

SensorFM was pretrained on about a trillion minutes of unlabeled sensor streams and then fine-tuned per endpoint. The architecture adapts a Vision Transformer, the model class that broke through in image recognition, to one-dimensional time-series signals from five sensors and 34 features that consumer smartwatches already expose. The playbook is the same pretrain-then-fine-tune pattern that delivered large language models, transplanted to continuous physiological signals.

The authors' framing in the paper and the blog post is that single-task supervised wearable AI stops scaling past about 35 endpoints, because every new condition demands its own labeled dataset. SensorFM's contribution, as they describe it, is treating smartwatch health as a foundation-model problem rather than a stack of narrow classifiers. One model absorbs the pretraining cost and amortizes it across every endpoint it is later asked to predict.

The framing has not been tested outside Google. The arXiv preprint has not been peer-reviewed, and the receipts include no third-party benchmark, no clinical validation, and no FDA submission. Coverage in TechTimes paraphrases the Google blog and the paper rather than introducing independent measurement. Readers who treat a smartwatch alert as a lab result are running ahead of what the public record supports.

The same caveat applies to two pieces of related Google Research wearable work that SensorFM is now grouped with. The Personal Health Agent is described in its accompanying blog post as a multi-agent research framework — combining data science, domain expert, and health coach roles — that the team has been building to work with these signals, designed to combine multiple foundation models behind a chat interface. LSM-2 is a foundation model for learning from incomplete wearable streams, the precondition for training across devices that are worn, charged, and forgotten in different patterns. Both are also Google-affiliated preprints or blog posts, with the same authorship pattern and the same missing third-party check.

An outside group would need a clinical dataset it did not give Google, a fixed evaluation protocol, and a paper that reports the numbers whether or not they match the Google blog post. None of that is in the current record. Cardiovascular risk is the endpoint most likely to be tested first, because cohort data already exists and the cost of a wrong call is concrete. Depression and anxiety are the ones where the gap between a wrist signal and a clinical diagnosis is largest, and where the claim deserves the most scrutiny.

If outside groups match the Google benchmark, the path from a wrist sensor to a clinical alert shortens considerably, and the cost of continuous population-scale screening drops with it. If they do not, and the held-out clinical numbers fall short, the field has just spent another year on a trillion-minute pretraining run that did not deliver the lab-test equivalence the blog post promised. Either outcome is a piece of evidence. The absence of either is not.

The next step is a download link, not a press cycle. SensorFM is on arXiv. An outside group that runs it on a held-out clinical dataset and publishes the numbers is the only thing that moves the claim from Google benchmark to settled result. Until that appears, the public record contains a foundation model that has only been benchmarked by its authors.