Executive Summary

Key Figures

These four numbers are the backbone of the investigation. The first is how much one company put into training. The middle two show how that practice was written down, or was not. The last belongs to a company that built the same kind of model without any consumer device.

Sources: the Science investigation, with tallies and context from Inside BCI and Tom's Hardware

The Science investigation was carried out over a year by Jonathan Moens, a freelance investigative reporter, working with the Pulitzer Center's AI Accountability Network. Its subject was the companies selling consumer EEG devices, the kind that measure the brain's electrical signals at the scalp. Fifty-seven of them were selected worldwide; the reporters read each privacy policy line by line and interviewed companies, experts and people who actually wear the devices.

The documents themselves were uneven. Four of the 57 had nothing that could be called a privacy policy. Eleven had one, but it covered only the information of website visitors and said nothing about the EEG the device and its app collect. Of the 42 that remained, roughly two-thirds stated plainly that they gather brain data.

The routes out of the company were wider than that. Every company with a policy but one permits sharing in de-identified or aggregated form, and only six have ruled out selling brain data. More than two-thirds state that if an acquisition or merger changes the owner, the data can change hands with it, without asking the user. Science added that at this stage, actually putting brain data up for sale still looks like an uncomfortable move for most of these companies.

Which box was empty is the finding of the investigation. Nine of the 57 had publicly said they were building large AI models of the brain, and at the outset not one had written that into its privacy policy. The model work itself was not hidden; some companies had announced it through press releases and other channels. It was absent only from the document a user reads and agrees to when signing up.

The diagram below lays out, from the top down, how much the 57 policies actually wrote. The top two rows are the cases where no document covers the EEG at all, and the further down a row sits, the more has been written. The bottom row has a different denominator. It counts only the nine companies that had disclosed brain-model work, and at the time of the review it was empty.

Owen Jones, who teaches law and biological sciences at Vanderbilt University, told Science what he made of that state of affairs: "That is not just deeply disturbing, it is cause for alarm and corrective action." Nor is there any guarantee that the nine who said they were building models are all of them. Science wrote that nine should be read as a floor, since others may be at work quietly.

Muse is a headband sold to help people meditate and to log their sleep. Electrodes at the forehead and behind the ears measure brain waves, so the device can chime when attention drifts or show the stages of last night's sleep in the morning. Interaxon confirmed to Science that it trained a brain foundation model on about 50,000 hours of the user recordings that had accumulated this way. Identifying details such as names and emails were removed before training, the company says.

A brain foundation model is a general-purpose model pretrained on large volumes of EEG and then reused across many functions, from scoring sleep stages to measuring focus. What Interaxon is after is a sharper reading of what happens in the brain during sleep and during concentration. At the time of reporting, the project was described as about halfway done.

The trouble is that none of this appeared anywhere in the privacy policy users agreed to when they signed up. The company amended the policy only after the reporter pointed that out. The current version contains a sentence saying de-identified data is used to train machine-learning models, foundation models included. It is written as a means of running the features Muse offers, such as sleep staging and brain health indicators, and it adds that identifying information is removed before anything goes into training.

Something did not make it into the amended policy. There is no condition requiring separate consent for use in model training. The company has also not said whether an unwilling user has a way out, or how many people's recordings went into the model.

This did not happen at one company alone. Emotiv, based in San Francisco, told Science it trains what it calls a large brain model on a portion of the recordings from users of its higher-end headsets, on public EEG datasets, and on material from neuromarketing research partnerships with universities and multinationals. It too amended its policy after the reporters made contact, and the wording it added differs from Interaxon's. The current version, effective 6 August 2026, says that pseudonymised neural data is used for training with the user's explicit opt-in consent. Two companies given the same prompt parted ways on how to handle consent.

Adnan Jinnah, the Muse user who appears in the story, has lived with insomnia for more than four years and has worn the band for thirty minutes before bed every night. He had no idea his brain recordings might also have trained an AI model. Told as much, what came back was closer to resignation than to anger: it was probably much like what big technology companies do with your data anyway.

Both companies amended their policies, which can make the matter look settled. What an amended policy promises, though, is how data will be handled from here. The data is already dissolved into the weights of a model that has finished training, and picking one user's recordings back out of it is effectively impossible. That is why Science put this point at the centre of the piece. The telling came afterwards; the training was already done.

Nita Farahany, professor of law and philosophy at Duke University, reads this as a problem of the context of consent. If someone agrees to one use and the actual use sits in an entirely different context, that agreement is not doing its job. Which is why, she argues, a company that wants to train AI on user data has to obtain "a specific and separate consent" for it. Agreeing to sleep tracking and agreeing to model training cannot be bundled into the same box.

On why brain waves should carry a stricter standard than that, Owen Jones, quoted earlier, put it this way: "Brain signals aren't heartbeats. Monitoring them is meaningfully different." His grounds were that a person is never the sum of the functions of their other organs, but can be described as the sum of their brain's activities.

The counterargument runs in the piece as well. Consumer devices are not as precise as clinical equipment, and EEG shows when activity happens far better than it pinpoints activity in deep structures such as the amygdala or the hippocampus. Nothing here reads thoughts today. Even so, most of the experts agreed that the right to know where your neural data goes is a separate question.

Nor is collecting data that way the only path to a model. Piramidal, a startup that appears in the same investigation, built its brain foundation model without touching recordings from consumer devices, using public EEG datasets and hospital partnerships instead. More than 4 million hours went into training. Far more data than Interaxon's, in other words, gathered through routes with a clear origin, which makes the argument that scale left no choice hard to sustain.

No US federal rule addresses neural data, brain waves included, head-on. The MIND Act (S.2925) was introduced in the Senate in September 2025, but what the bill does stops at directing the Federal Trade Commission to study neural data for a year and then recommend a regulatory framework. It sets no standard to follow now, and there has been no visible movement since it was introduced.

The states are where the movement is. California classifies neural data as sensitive personal information, Colorado folds it into the biometric category, and Montana covers information from both the central and the peripheral nervous system. Connecticut has gone furthest. Its rule, in force since 1 July 2026, requires explicit consent before a resident's neural data is processed, and a second, separate consent before that data is handed to a third party for payment. It carves out no exemption for company size or user count.

The diagram below lays out, along a single line, how far apart the standards sit for the same data depending on where you are. The left end is where no standard exists; the right end is where a state body looks at the data regardless of what the person agreed to.

Outside the United States the spread grows wider. Chile was the first country to write neurorights into its constitution, in 2021, and in 2023 its supreme court ordered Emotiv to delete a former senator's brain data. At the other end, companies in China are collecting brain waves from truck drivers and construction workers, in some cases without anonymisation. About 2,000 EEG safety helmets are in use on building sites in Guangzhou, Chongqing and Beijing, and the recordings accumulated in the name of watching fatigue and attention run past 300,000 hours. Another company has opened drivers' brain waves to live viewing by traffic police. UNESCO and the OECD have both recently flagged the spread of consumer brain-computer interfaces as a privacy question.

The question of where training data came from has been put to text and images for several years now. What this investigation shows is that the same question has reached signals that come out of the body. Text can be deleted and written again, images can be walled off from download, but brain waves measured while someone was asleep are not the sort of thing you can swap out. The words "AI training" were missing from the consent document, and were filled in only once a reporter asked.

When people judge whether data is fit for model training, what they usually look at is volume and quality. This episode pulls a third item forward: under what promise was the data gathered. Data with a broken chain of consent does not become usable just because 50,000 hours of it have piled up. The counterexample sits inside the same story. Piramidal gathered 4 million hours through routes with a clear origin.

For anyone already training models on user data, or planning to, four things can be confirmed on paper.

• Does the notice given at collection time state the training use? Is "improving the service" being treated as cover for model training?
• When the uses grow, is there a route back to the user for consent? Is a notice of a policy amendment standing in for that?
• Is there a record of which training run took in data under which consent terms? After a model has split into several versions, can that record still be traced back?
• When a user asks for deletion, does the request stop at the store or does it reach the model? If it does not reach, how will that be explained to the user?

None of the four asks whether the data is accurate. They ask under what promise it arrived. A policy can be amended in a day and a trained model cannot, and that is the most practical lesson this investigation leaves behind.

Thank you for reading this far. The investigation itself is at Science, and Inside BCI has set out the policy tallies and the regulatory context separately. If you have had to write the consent language for using user data in model training, or to decide how far that language should go, we would be glad to hear about it.

The Investigation and Its Summaries

• 1.Moens, J. (2026). "Wearable device makers are harvesting brain data to train AI—without telling users." Science. (Reported with the Pulitzer Center's AI Accountability Network)
• 2.Inside BCI. (2026). "Brain-wearable makers left AI training on users' brain data out of their privacy policies, Science finds."
• 3.AI Weekly. (2026). "Consumer EEG Headband Makers Have Quietly Harvested Tens of Millions of Hours of Brain Data for AI Training."

Industry and Press

• 4.Tom's Hardware. (2026). "Wearable sleep trackers are using your brainwaves to train AI models without explicit user consent."
• 5.FourWeekMBA. (2026). "Muse Headband Maker Says It Trained AI on 50,000 EEG Hours."

Regulation and Law

• 6.Tech Times. (2026). "Connecticut Classifies Brain Data as Sensitive Starting July 1: New Rights for EEG Wearable Users."