A year-long investigation by Science and the Pulitzer Center's AI Accountability Network found that consumer sleep and meditation wearables supplied electroencephalography (EEG) recordings for artificial intelligence training.

The investigation identified nine companies building brain foundation models, including Interaxon, maker of the Muse headband, and Emotiv. It found that their earlier privacy policies did not clearly disclose that recordings could be used to train AI models.

The companies had collected tens of millions of hours of EEG data, the investigation found. That figure is not a verified count of recordings used to train AI, and the findings do not establish that every company or customer recording was involved.

EEG measures electrical signals from the brain. Sleep and meditation wearables use these recordings to provide information about users' activity, while AI developers can use them to train models to identify patterns in brain signals.

What Interaxon Disclosed About Sleep Tracker Data

Interaxon, the Canadian company behind the Muse sleep and meditation headband, told Science it had trained an AI model using about 50,000 hours of EEG recordings, most collected from device users. The figure was provided by the company and has not been independently audited.

Interaxon said it removed direct identifiers, including names and email addresses, before using recordings to develop the model. Its updated privacy policy says de-identified session data may be used to improve algorithms and machine-learning models, including foundation models.

The investigation found that the previous policy did not explicitly mention this use.

The company updated its policy after Science asked about its data practices. Interaxon chief marketing officer Nadia Kumentas told the journal: 'We aim to use as much of that eligible corpus as possible.'

Read more

Emotiv Also Revised Its Privacy Policy

Emotiv was using customer EEG recordings to train a model called Large Brain, according to the investigation. After Science contacted the company, Emotiv updated its privacy policy to disclose the use of de-identified neural data for model training, subject to opt-in consent.

The investigation reviewed privacy practices at dozens of consumer companies and identified nine that had publicly said they were developing brain foundation models. It found that none had clearly disclosed this specific use to consumers before the inquiry.

Foundation models are trained on large datasets and can be adapted for particular tasks. The models examined in the investigation learn patterns in neural signals and may be adapted for uses such as identifying sleep stages.

What the Investigation Cannot Establish

Both companies have revised their policies, but questions remain about data already collected. Emotiv's terms say withdrawing consent stops future data collection, but does not remove contributions already incorporated into trained model weights. The available information does not establish whether the same applies to Interaxon.

Nor did the investigation establish whether de-identified EEG recordings could be reidentified. Removing names and email addresses reduces direct identifiers, but it is not the same as obtaining informed, specific consent for AI training.

The investigation made no legal determination about either company's compliance with privacy law. That would depend on the jurisdiction, the wording of each policy and how consent was sought. For customers, the distinction is between agreeing to a device's sleep or meditation function and being clearly informed that its recordings may also help train AI.

Read more