Brain-to-brain interfaces

Brain-to-brain interfaces are systems that connect two nervous systems by decoding a signal from one brain and delivering stimulation to another, without the sender speaking or the receiver listening. Every demonstration so far chains together two established technologies: a Brain–computer interface that reads out a single variable, and a neurostimulation device that writes a crude signal in. Nothing about the arrangement is telepathic in the sense the popular coverage implies.

The achieved information rates are the essential fact. Across the published human experiments, transmission has amounted to roughly one bit per trial, with trials taking tens of seconds — a few bits per minute, against roughly forty bits per second for ordinary speech. What is transmitted is not a thought but a pre-agreed binary code that the experimenters defined in advance.

§How it works

The sender's side is a decoder. In the non-invasive human experiments the sender looks at one of two flickering targets, producing a steady-state visual evoked potential at the corresponding frequency in occipital cortex, which scalp electroencephalography can classify reliably. Alternatively the sender imagines moving a hand, producing a detectable change in sensorimotor rhythms. Either way, the output is a single binary decision — the simplest possible case of Neural decoding.

The receiver's side is a stimulator, drawn from the standard toolkit of Non-invasive neuromodulation. Transcranial magnetic stimulation over primary motor cortex can evoke an involuntary hand movement; over occipital cortex it can evoke a phosphene, a brief spot of perceived light. The bit is encoded as stimulate or do not stimulate. In animal work the write channel has been intracortical microstimulation through an implanted array, or focused ultrasound delivered through the skull.

What passes between the brains is therefore a bit chosen by the experimental protocol. The receiver does not experience the sender's percept; the receiver sees a flash, or feels a hand move, and applies a rule learned beforehand about what a flash means.

§Development history

1. 2013
   Rat to rat
   A Duke group links an encoder rat's motor cortex to a decoder rat's via microstimulation; the decoder rat performs a lever task above chance using information it could not otherwise obtain.
2. 2013
   Human to rat
   A Harvard-affiliated team uses human EEG to trigger focused ultrasound over a rat's motor cortex, producing a tail movement.
3. 2013–2014
   Human to human
   A University of Washington pair transmit a motor-imagery signal from one person's EEG to magnetic stimulation over another's motor cortex, moving a finger to fire in a video game.
4. 2014
   Words across continents
   A team encodes 'hola' and 'ciao' in binary, sending them from an EEG subject in India to subjects in France who perceive the bits as phosphenes.
5. 2015
   Multi-animal networks
   Nicolelis's group links several rat and several monkey brains into 'brainets' that jointly perform computations or control a virtual arm.
6. 2019
   BrainNet
   Three people collaborate on a Tetris-like task, two sending rotate-or-not decisions by EEG and the third receiving them as phosphenes, with accuracy around 80 percent.

The 2013 rat experiment remains the most substantive result, because it involved implanted electrodes on both ends and the receiving animal genuinely improved at a task using transmitted information.¹ The human work has all been non-invasive, and therefore restricted to the coarsest possible read and write.²³⁴ BrainNet, which linked three people in a shared task, is the largest network demonstrated and still transmitted a single bit per sender per turn.⁵ Multi-animal "brainets" have shown that several brains can be pooled to control one output, which is a different and somewhat less surprising result: it is population decoding across skulls.⁶

Why "telepathy" is the wrong word

None of these systems transmits meaning. The bit is meaningful only because both participants were told in advance what it would signify. A system with the same architecture and a light bulb instead of magnetic stimulation would work identically, faster, and with less equipment. The neuroscientific content of the demonstrations is that a decoded signal can be delivered as a percept, not that brains can share content.

§Why the bandwidth is so low

Three obstacles compound, and only one of them is an engineering problem.

Reading. Non-invasive read-out through the skull resolves centimetre-scale cortical activity at best, which caps the sender's channel at a few bits per trial. Implanted electrodes would raise this substantially, as the results from Utah array and Electrocorticography interfaces work show, but nobody has implanted a healthy volunteer to serve as a sender.

Writing. This is the harder half. Transcranial magnetic stimulation activates on the order of a cubic centimetre of cortex indiscriminately; focused ultrasound is better localized but still addresses large populations; intracortical microstimulation is finer but still activates hundreds of cells in a pattern nothing like natural activity. The read-out fidelity of current neurotechnology far exceeds its write-in fidelity, which is also why sensory feedback in Neuroprosthetics lags motor decoding. Precise write-in through optical control is possible in animals via Optogenetics, but requires genetically modifying the target neurons.

No common code. Even with perfect read and write, the deeper problem is that two brains do not share a representational format. The population activity encoding a particular concept in one person's cortex has no fixed correspondence to activity in another's; representations are shaped by individual developmental and learning history. Delivering A's neural pattern to B would not reproduce A's experience in B, any more than installing one computer's memory contents in another with a different architecture would run the program.

The plausible workaround is learning rather than translation. A receiver could, in principle, learn to interpret an arbitrary but consistent stimulation code, in the way that users of Sensory augmentation devices come to experience a substituted signal as a perception rather than a puzzle. That has not been attempted at any scale for brain-to-brain transmission, and it would make the channel a new sense to be learned rather than a shortcut past learning.

§Ethics

The experiments raise a problem unusual in neurotechnology: the receiver's body moves without the receiver intending it. Participants in the human studies consented to precisely that, but the arrangement makes the question of agency concrete in a way that a cursor decoder does not. Related concerns about who holds the transmitted signal, and what could be inferred from it, fall under Mental privacy and the emerging Neurorights proposals.

There is also a persistent framing problem. Commercial neurotechnology has borrowed the vocabulary — Neuralink markets its cursor-control product under the name Telepathy — in ways that inflate public expectations of what neural interfaces do. The distance between typing with a decoder and sharing an experience is not a matter of degree.

§Outlook

No path from the current demonstrations to meaningful brain-to-brain communication has been articulated in technical detail. The requirements are a high-resolution write channel that does not exist, a solution to the shared-code problem that no one has proposed beyond "the receiver learns", and an ethical case for implanting healthy people that nothing currently justifies.

The idea nonetheless persists because it sits at the centre of several other speculative programmes: proposals for Human–AI merger often assume a rich bidirectional neural channel, scenarios built on Whole brain emulation assume that a mind's content can be extracted and moved, and arguments about merged or networked minds in the Mind uploading literature take shared access between substrates for granted. Those arguments generally inherit the assumption that transmitting neural activity transmits its content. The brain-to-brain experiments are the closest thing to a direct test of that assumption, and what they show is a single bit crossing a gap that everything else about the two brains has to be arranged in advance to interpret.

§See also

• Brain–computer interface
• Neural decoding
• Non-invasive neuromodulation
• Sensory augmentation
• Mental privacy
• Human–AI merger
• Neuroprosthetics
• Mind uploading

§References