Reading the mind

What brain-computer interfaces actually do, and what is being promised for them

A Few technologies arrive with as much accompanying noise as this one, and few are as poorly served by it. The idea of connecting a brain directly to a machine has been established in principle for half a century. Electrodes placed on the surface of the cortex or inserted into it record the electrical activity of populations of neurons; software learns which patterns accompany which intentions; and the resulting signal is used to move a cursor, a robotic arm or a wheelchair. The demonstrations are genuine, the participants are real people with severe paralysis, and the field has produced results that would have seemed fantastic in the 1990s.

B What the demonstrations conceal is the distance between a laboratory result and a device somebody can live with. Implanted electrode arrays provoke a response from the immune system: scar tissue forms around them, the signal degrades over months and years, and an array that performs superbly in the first weeks may be useless within a few years. The connection to the outside world has often been a socket in the skull, with the infection risk that implies. Recording sessions are typically supervised by technicians who recalibrate the system as the signals drift. None of these are objections to the concept, but they are the reasons the number of people living with such a device remains in the dozens. Fully implanted systems that transmit without a wire through the skin are now in trials and remove one of these obstacles; they do not remove the others, and the scar tissue in particular is a response to the presence of the electrode rather than to anything that can be engineered away.

C The alternative is to record from outside the head, which is safe, cheap and severely limited. Electrical activity measured at the scalp is blurred by the skull and the tissue between, so what arrives is the summed activity of very large populations, filtered and smeared. Useful signals can be extracted, and non-invasive systems drive some assistive devices, but they cannot approach the precision of an implant for the same reason that a microphone outside a stadium cannot pick out a single conversation. Every advance in machine learning applied to the problem has improved matters somewhat and has not changed this fundamental limit.

D The most impressive recent work concerns speech. Participants who cannot move or speak have had their attempted speech decoded from motor cortex activity and rendered as text or synthesised voice, at rates that have risen from a few words per minute to something approaching conversational speed. The achievement is real and its significance for people who have lost the ability to communicate is difficult to overstate: a person able to produce a sentence a minute has a life materially different from one able to produce a word. It is also, at present, the product of intensive individual calibration: a system trained on one participant does not transfer to another, and the training is measured in months.

E Commercial attention has arrived in force, and with it a familiar difficulty. Companies pursuing implanted interfaces have attracted very large investments on the strength of demonstrations, and their public statements have described applications — memory enhancement, direct communication between people, control of general-purpose computers by thought — that bear no relation to what any laboratory can currently do. The gap between the clinical literature and the promotional material is now wide enough that clinicians working in the field have begun publicly correcting it, which is not a healthy sign for a discipline that depends on the confidence of desperate patients.

F The ethical questions that follow are not the ones usually raised. Public discussion turns on mind-reading, which the technology cannot do and will not do soon; the immediate problems are more mundane and better documented. A person whose implant is controlled by adaptive software may find it difficult to say whether an action was theirs, which matters for responsibility. Neural recordings are data, and no regulatory framework treats them differently from any other data a company holds. Most concretely, participants in trials of retinal implants have been left with obsolete hardware in their bodies when the manufacturers ceased trading, unable to obtain support for a device they cannot easily have removed, and there is no reason to think a cortical implant would fare better.

G None of this argues against the work. A technology that restores communication to someone who has lost it needs no further justification, and the trajectory of the last decade is genuinely remarkable. But the field would be better served by a description of itself that matched the evidence: a small number of participants, devices that degrade, results that do not yet transfer between individuals, and a set of governance questions that are being deferred while the promises accelerate. The honest version is impressive enough, and it has the additional merit of being what will actually happen.