Google's new AI can read your body fat from two phone photos — and it beats your smartwatch

PhotoScan is still a research prototype, but its accuracy already lands closer to a real DXA scan than the sensor in your smartwatch. Aliteq Lena Fischer · AI & Local Compute Editor

PhotoScan is an experimental Google Research AI model that estimates body fat percentage, fat distribution (android/gynoid ratio) and visceral-vs-subcutaneous fat from two smartphone photos — one…

It was pre-trained on 35,323 UK Biobank records with MRI and DXA scan data, then fine-tuned on a 677-person cohort and validated on an independent 132-person group.

Its body-fat-percentage error (mean absolute error of 2.13-2.15 percentage points) beats the bioelectrical impedance sensors used in smartwatches, per Google's own comparison.

It can also flag insulin resistance risk from the same two photos, with 76% classification accuracy (AUROC 0.760) — close to a clinical DXA scan combined with demographic data (AUROC 0.773).

It's a research prototype, not a shipping Pixel or Fit feature — Google hasn't announced when or if it becomes a real product.

The accuracy here is genuinely impressive, but an AI trained to read your body composition from photos raises exactly the kind of health-data question Google has stumbled on before. Before this…

Read the full story

Aliteq Google's new AI can read your body fat from two phone photos — and it beats your smartwatch Read the full story on Aliteq