A new study pokes a hole in the idea that one AI model can learn to read anyone's emotions from a smartwatch.
Researchers built RL-PPG, a self-supervised model trained on unstructured, real-life photoplethysmography data - the heart-rate signal your wearable already tracks. As a sanity check, they tested it on physical activity recognition and got a nearly 5x jump in accuracy over baseline models. But when they pointed the same general model at detecting intense emotions in real-life settings, using a leave-one-subject-out test, it could not beat naive baselines. Switching to an across-time validation approach, the team found that fine-tuning on an individual's own data mattered far more than any population-level pretraining.
The gap matters because wearable makers increasingly pitch AI wellness and mood-tracking features as a selling point. This research suggests those features work fine for objective measures like steps or workouts, but fall apart for anything as subjective as emotion, unless the model learns your specific physiology first. That is a much harder, slower, and more data-hungry product to build than a one-size-fits-all model.
It is a useful reality check against marketing copy that treats AI-knows-how-you-feel as a solved problem. The researchers released their code and pretrained weights, effectively challenging anyone building emotion-detection wearables to prove personalization actually happens before buying the claim.