← Latest articles

Companies Use Wearable Sleep Trackers to Harvest Brain Data for AI Training

File photograph: OURA Ring - 2.
Photo credits
Photo: Kyu3a · CC BY-SA 4.0 · Cropped within the original photograph to 16:9, resized, and converted to WebP; no padding or synthetic background.

Manufacturers of consumer sleep headbands are turning customer brain-wave recordings into training material for artificial intelligence models, often without explicitly disclosing the practice in their privacy policies. An investigation by Science and the Pulitzer Center's AI Accountability Network found that neurology companies have amassed tens of millions of hours of electroencephalography data.

Interaxon, the Canadian firm that manufactures the Muse sleep and meditation headband, is among the companies utilizing this data. The devices monitor electrical brain activity alongside other physiological measurements while users sleep to generate rest-tracking reports. However, Interaxon has also used approximately 50,000 hours of those brain-wave recordings to train an artificial intelligence model. Nadia Kumentas, the company's chief marketing officer, stated that the firm aims to use as much of that eligible data corpus as possible.

To address privacy concerns, Interaxon removes identifying details such as names, email addresses, and home addresses before processing the recordings. Unlike systems trained on large volumes of written text, these brain foundation models are supplied with recordings of electrical brain activity. The models learn recurring patterns that developers can adapt to recognize different brain states or assist with specific research tasks.

Other organizations are also accumulating neural recordings on a large scale. The AI startup Piramidal partnered with the Cleveland Clinic to train its model on more than 4 million hours of brain-wave data, developing a seizure-detection system that correctly identified nine out of ten seizures in a pilot study. In San Francisco, startup Coundit built a custom headset to capture neural activity alongside speech and text to create translation software. Meanwhile, Meta developed a non-invasive interface called BrainQwerty to translate brain activity into keyboard inputs, improving its average word accuracy from 40% to 61% in a second-generation system.

Despite these capabilities, consumer devices cannot read a person's thoughts. Most wearable EEG headbands use between one and six electrodes, which is considerably fewer than the 19 to 25 sensors found in clinical systems, yielding less comprehensive recordings.

Separate software vulnerabilities have also exposed sensitive neural information to unauthorized users. An engineer who purchased a smart sleep mask from a small research company encountered connection issues and investigated the device's mobile application. Using decompilation tools, he discovered access credentials hard-coded into the app binary and shared across devices. By connecting to remote data endpoints using those credentials, he was able to view real-time brain-wave readings from other users around the world.

Summary

Consumers purchase wearable headbands to monitor their rest, unaware that companies may harvest their neural recordings to train artificial intelligence models. Concurrently, software security flaws in some connected devices leave sensitive brain data exposed to strangers, underscoring ongoing oversight and privacy challenges in the consumer neurotechnology market.