A brain signal can now be collected by a medical device, a brain-computer interface or, in simpler forms, wearable technology. This is not mind-reading in the science-fiction sense. It is the measurement of electrical or physiological activity from which a system tries to recognise patterns. When artificial intelligence enters this process, the brain no longer produces only clinical data. It becomes a source of potentially reusable inferences about health, attention, intention and behaviour.
On September 8, the European Commission published Governing Neuro-AI, an opinion by the European Group on Ethics in Science and New Technologies. It describes the convergence of neuroscience and AI as a change of scale: assessing one implant or application is no longer enough. We must examine the whole infrastructure of sensors, platforms, clouds, archives, models and organisations able to turn a biological signal into a decision.
The crucial distinction is between neurodata and inference. Raw data may be a sequence of impulses; an inference is what a model claims it can deduce from that sequence. A prediction about attention, disease risk or emotional response may be more sensitive than the original signal. Protecting only the collected file is therefore insufficient. The interpretations produced, stored, sold or connected to other profiles must also be governed.
Brain Foundation Models make the issue more concrete. As large language models learn structures from vast amounts of text, these systems search for regularities in brain data from many people, devices and tasks. They may support research, diagnosis, rehabilitation and new forms of communication for people unable to speak or move. Yet a general model can be adapted to purposes far removed from those for which the data were collected, making one-time consent inadequate.
Consent remains essential, but it cannot carry the full burden of protection. A person may approve the use of their data for research without being able to predict what a model will do five years later. Purpose limitations, genuine withdrawal, access logs and rules preventing clinical participation from becoming an unlimited commercial resource are also required. The more powerful reuse becomes, the more specific and intelligible authorisation must remain.
High-risk applications are not only medical. An employer may want to measure attention or fatigue; an insurer may search for vulnerability; a school may classify concentration and learning; an authority may attach meaning to physiological reactions. Even when a prediction is statistically accurate, its legitimacy is not automatic. Democratic societies can decide that some dimensions of a person should not become measurable by whoever holds more power.
Mental privacy is therefore more than secrecy. It includes freedom from being reduced to a prediction, from decisions based on presumed internal states and the right to challenge what a system claims to have recognised. A model may be wrong, but it may also create a label so plausible that it shapes the behaviour of doctors, employers or institutions before the person can dispute it.
The question is industrial too. Whoever controls large neurodata archives and the capacity to train cross-purpose models can define standards, services and conditions of access for research and healthcare. The European opinion proposes treating Neuro-AI as public-interest infrastructure, strengthening independent capacity, democratic oversight and international cooperation. Sovereignty does not mean closing data; it means ensuring that its collective value is not governed solely by a few providers.
None of this makes every neurotechnology dystopian. Well-designed interfaces can restore communication and autonomy, support rehabilitation and reveal clinical signals that were previously difficult to interpret. Precisely because the benefits are real, rules should do more than prohibit. They must create trustworthy conditions for research, separate care from surveillance and require transparency about performance, uncertainty, training populations and error.
When the brain becomes an infrastructure, the boundary between body and data stops being purely individual. Every model learns from many people and may affect many others. The question is not only who owns a recording, but who may turn it into knowledge, who has the right to apply that knowledge and who answers for the consequences. Neuro-AI forces us to establish one essential rule in advance: the technical ability to infer something from the mind does not automatically create the right to do so.