The electroencephalography recordings gathered by consumer sleep headbands have quietly become a training corpus. An investigation published in Science, conducted with the Pulitzer Center's AI Accountability Network, found that the companies behind these wearables hold tens of millions of hours of EEG data, and that several of them are using the recordings to develop brain foundation models, in some cases before their privacy policies said anything about that use.

What Interaxon did with 50,000 hours

Interaxon, the Canadian company behind the Muse sleep and meditation headband, has already trained an AI model on about 50,000 hours of EEG recordings gathered from customers whose devices monitor brain activity as they sleep. Chief marketing officer Nadia Kumentas said the company strips names, emails and addresses before training, and that it aims "to use as much of that eligible corpus as possible." Before Science asked questions, Interaxon's privacy policy did not spell out that brain recordings could feed foundation models. It has since been updated, and Kumentas acknowledged that "users should be explicitly informed." For customers whose recordings were already used, the disclosure came late.

Emotiv, another EEG wearable maker, trains its Large Brain model partly on recordings from customers using its high-end devices and also revised its policy after the investigation, moving to opt-in consent for de-identified neural data. The pattern stretches wider than headbands: data brokers have been buying work email inboxes for AI training, and brain recordings are simply the newest category of personal signal to be repurposed this way.

The research side of this is genuinely promising. Piramidal, an AI startup, trained a model on more than 4 million hours of EEG data in partnership with Cleveland Clinic and built a seizure-detection system that caught nine out of ten seizures in a pilot study after fine-tuning on the clinic's recordings. But consumer hardware is a different instrument: wearable EEG devices carry one to six electrodes, against the 19 to 25 found in clinical systems, and their recordings are far less comprehensive.

The governance is where the gap sits. Science found that more than two-thirds of the companies studied explicitly allow transfers of user data to corporate acquirers without fresh consent, which matters because a small neurotechnology firm's accumulated recordings can change hands in an acquisition. Colorado and California have amended their consumer privacy laws to cover neural data; most other states have not. Earlier this year an engineer who dug into a smart sleep mask he owned found hardcoded credentials shared across devices and, through them, access to other users' brainwave recordings.

A model that reads signals, not thoughts

These brain foundation models differ from language models in what they consume: instead of text, developers feed in electrical recordings and the model learns recurring patterns that can be adapted to recognize brain states or assist research. Nothing on the market today reads thoughts. What the Science investigation documents is a consent problem, not a capability problem, and the field has already produced stranger headlines than a sleep headband: researchers put a simulated fruit fly brain to work mining bitcoin just to show the model could drive behavior. The recording infrastructure, in other words, is being built faster than the rules that govern it, and the first wave of policy updates has arrived only after journalists asked.

Source: Tom's Hardware