Data from our brains is becoming a resource for artificial intelligence

A headband worn to improve sleep, aid meditation or measure concentration records the brain’s electrical activity and displays the results on a mobile phone. Those who buy it believe they have purchased a service. Instead, without having been adequately informed, they may have contributed to building an industrial-scale repository of brain data intended for training artificial intelligence. This is documented in a year-long investigation by Jonathan Moens, published in Science. The story reveals just how dangerous it is to entrust the market – without effective public oversight – with the collection and exploitation of information that comes directly from the organ responsible for perceptions, emotions, memory and thought, and about which we do not yet know how much we will eventually be able to learn through developments in artificial intelligence.

These devices use electroencephalography, or EEG: electrodes on the scalp record variations in electrical potential produced by the activity of large populations of neurons. Commercial devices provide signals that are far less detailed than those from clinical equipment and do not freely read thoughts or memories. However, they can feed into algorithms capable of recognising sleep patterns or estimating certain cognitive states. The sensitivity of these recordings also depends on what future algorithms will be able to extract from them.

Moens examined the privacy policies of 57 companies marketing these neurotechnologies. Four had none at all, and eleven only provided statements relating to website visitors, without explaining how brain signals were processed. Among the companies with privacy policies, all but one permitted some form of sharing of aggregated data or data stripped of personal identifiers with third parties, without always clarifying whether EEG recordings were included. Only six expressly ruled out the sale of brain data. This does not prove that actual sales took place, but it does illustrate just how unclear the fate of such sensitive information was.

Interaxon, the Canadian manufacturer of the Muse headband, told Science that it had trained a model using approximately 50,000 hours of EEG data, most of which came from users of the product. According to the company, names and email addresses had been removed from the recordings. However, the privacy notice did not explicitly state that the signals would be used to train a general model of the brain. Following questions from the journalist, Interaxon amended it, acknowledging the need to inform users. A key aspect of the commercial agreement was only clarified following an independent investigation, after the training had already taken place.

Emotiv has also explained that it uses, for its Large Brain Model, some of the recordings from users of advanced devices, alongside other sources. Following enquiries from Science, it updated its terms and conditions to require explicit consent for training with pseudonymised neural data. This change confirms that using a device and contributing to the development of a commercial AI product are distinct purposes. Users must be able to authorise one without being obliged to accept the other.

To understand the economic value of the operation, one must consider the brain’s ‘foundational models’. These are systems trained on large quantities of data to learn statistical regularities and then adapted to specific problems, ranging from the classification of sleep stages to the recognition of certain neurological abnormalities. The competitive advantage also stems from the exclusive possession of signal collections that are difficult to reproduce. Manufacturers can build these collections using devices paid for by consumers, who are often unaware that their daily use contributes to creating a transferable and monetisable industrial resource.

Removing the name from the file does not automatically resolve the issue. Pseudonymisation means separating data from direct identifiers; anonymisation means making it not reasonably possible to trace the data back to the individual. These are different conditions. EEG tracings contain individual characteristics that are also studied as potential biometric identification tools. An algorithm can also derive information about cognitive states and other sensitive characteristics. Patrick Magee, Marcello Ienca and Nita Farahany explained in Neuron that mental privacy requires consideration not only of the original recordings but also of the inferences drawn from the data. Simply removing first and last names therefore does not provide sufficient safeguards.

There is also a technical asymmetry. Deleting a record from an archive is a traceable operation; removing the contribution it has made to the training of a neural network is far more difficult and not always verifiable. The right to withdraw consent is therefore at risk of being exercised only after the company has already incorporated the results of data usage into the product. Anyone who stops using the device may still have contributed to a model that remains usable.

The problem arises in an even more serious form in the workplace. A second report in Science details the use of helmets and other EEG devices to detect fatigue among workers and drivers, particularly in China. Preventing accidents is a legitimate aim, but the consent given by an employee subject to workplace surveillance lacks the freedom of that of a consumer who can choose not to use such a device. If the recording is linked to the worker’s identity and passed on to superiors or the authorities, the assessment of fatigue can turn into constant monitoring of the individual. Even errors in the algorithm can have professional consequences.

Europe has the General Data Protection Regulation, which requires specific purposes, transparency and special safeguards for health and biometric data where the conditions laid down by law are met. Not every EEG automatically falls within the same legal category, and the issue requires specific scrutiny of data collection and reuse practices. In the United States, Colorado and California have already extended legislative protections for neural data. General protections therefore require specific application.

The authorities should require a verifiable description of the data stored and its intended uses; separate, revocable consent that is not a prerequisite for using basic functions when commercial training is involved; and independent audits of training databases and transfers between companies. Stricter limits are needed in the workplace, as the right to opt out carries little weight when economic dependence comes into play. The origin of the data used by the models must also be verifiable.

The crucial issue is the distribution of power. Companies can transform data collected for a specific purpose into strategic assets, which can be exploited even when the company changes hands, whilst those who generated the data remain unaware of the full extent of the operation. The market rewards the accumulation of data and the success of models, but does not automatically account for the cost of future intrusions into individuals’ private lives. Entrusting companies with the task of defining the limits of what they can collect amounts to allowing those who profit from this expansion to determine the rules. Brain research can be carried out using legally collected data and independent oversight. What Science has documented, however, is a removal of decision-making power from those concerned, made possible by the weakness of public oversight. And it is precisely this that cannot be left to the market.