Brain-to-brain interfaces: revision 1 to 2

Revision 1

Revision 2

categories: ["cybernetics", "minds"]

tags: ["bci", "telepathy", "tms", "neural decoding", "neurostimulation", "bandwidth"]

summary: "Systems that link one nervous system to another by decoding activity from a sender and delivering stimulation to a receiver, so far at rates of a few bits per minute."

summary: "Systems that link one nervous system to another by decoding activity from a sender and delivering stimulation to a receiver, so far at rates of a few bits per minute."

updated: "2026-07-27"

updated: "2026-10-29"

humanEvidence: "All human demonstrations are non-invasive: volunteers have exchanged roughly one bit per trial through scalp EEG and magnetic stimulation, and no healthy person has been implanted for this purpose."

humanEvidence: "All human demonstrations are non-invasive: volunteers have exchanged roughly one bit per trial through scalp EEG and magnetic stimulation, and no healthy person has been implanted for this purpose."

access: "Nothing to obtain: the systems exist only as laboratory setups assembled from research EEG and stimulation equipment, with no product or service on offer."

access: "Nothing to obtain: the systems exist only as laboratory setups assembled from research EEG and stimulation equipment, with no product or service on offer."

reversibility: "reversible"

"caption": "Experimental neurotechnology",

"rows": [

{ "label": "First animal demonstration", "value": "2013 (rat to rat)" },

{ "label": "First human demonstration", "value": "2013–2014" },

{ "label": "First human demonstration", "value": "2013-2014" },

{ "label": "Read method", "value": "EEG or implanted electrodes" },

{ "label": "Write method", "value": "TMS, ultrasound, microstimulation" },

{ "label": "Demonstrated rate", "value": "A few bits per minute" },

}

```

**Brain-to-brain interfaces** are systems that connect two nervous systems by decoding a signal

**Brain-to-brain interfaces** are systems that connect two nervous systems by decoding a signal from

from one brain and delivering stimulation to another, without the sender speaking or the receiver

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

writes a crude signal in. The arrangement is not telepathic in the sense popular coverage implies.

coverage implies.

The achieved information rates are the essential fact. Across the published human experiments,

The information rates achieved are the central fact. Across the published human experiments,

transmission has amounted to roughly one bit per trial, with trials taking tens of seconds — a few

transmission has amounted to roughly one bit per trial, with trials taking tens of seconds. That

bits per minute, against roughly forty bits per second for ordinary speech. What is transmitted is

works out to a few bits per minute, against roughly forty bits per second for ordinary speech. The

not a thought but a pre-agreed binary code that the experimenters defined in advance.

transmitted content is a binary code the experimenters defined in advance, not a thought.

## 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

flickering targets, which produces a steady-state visual evoked potential at the corresponding

in occipital cortex, which scalp electroencephalography can classify reliably. Alternatively the

frequency in occipital cortex, and scalp electroencephalography can classify that reliably.

sender imagines moving a hand, producing a detectable change in sensorimotor rhythms. Either way,

Alternatively the sender imagines moving a hand, producing a detectable change in sensorimotor

the output is a single binary decision — the simplest possible case of [[neural-decoding]].

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

perceived light. The bit is encoded as stimulate or do not stimulate. In animal work the write

write channel has been intracortical microstimulation through an implanted array, or focused

channel has been intracortical microstimulation through an implanted array, or focused ultrasound

ultrasound delivered through the skull.

delivered through the skull.

What passes between the brains is therefore a bit chosen by the experimental protocol. The

What passes between the brains is therefore a bit chosen by the experimental protocol. The receiver

receiver does not experience the sender's percept; the receiver sees a flash, or feels a hand move,

does not experience the sender's percept. The receiver sees a flash, or feels a hand move, and

and applies a rule learned beforehand about what a flash means.

applies a rule learned beforehand about what a flash means.

## Development history

[

  { "year": "2013", "title": "Rat to rat", "text": "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." },

  { "year": "2013", "title": "Rat to rat", "text": "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." },

  { "year": "2013", "title": "Human to rat", "text": "A Harvard-affiliated team uses human EEG to trigger focused ultrasound over a rat's motor cortex, producing a tail movement." },

  { "year": "2013", "title": "Human to rat", "text": "A Harvard-affiliated team uses human EEG to trigger focused ultrasound over a rat's motor cortex, producing a tail movement." },

  { "year": "2013–2014", "title": "Human to human", "text": "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." },

  { "year": "2013-2014", "title": "Human to human", "text": "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." },

  { "year": "2014", "title": "Words across continents", "text": "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." },

  { "year": "2014", "title": "Words across continents", "text": "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." },

  { "year": "2015", "title": "Multi-animal networks", "text": "Nicolelis's group links several rat and several monkey brains into 'brainets' that jointly perform computations or control a virtual arm." },

  { "year": "2015", "title": "Multi-animal networks", "text": "Nicolelis's group links several rat and several monkey brains into 'brainets' that jointly perform computations or control a virtual arm." },

  { "year": "2019", "title": "BrainNet", "text": "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." }

  { "year": "2019", "title": "BrainNet", "text": "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

The 2013 rat experiment remains the most substantive result, because it used implanted electrodes at

electrodes on both ends and the receiving animal genuinely improved at a task using transmitted

both ends and the receiving animal genuinely improved at a task using transmitted

information.[^paisvieira2013] The human work has all been non-invasive, and therefore restricted to

information.[^paisvieira2013] The human work has all been non-invasive, and is therefore restricted

the coarsest possible read and write.[^yoo2013][^rao2014][^grau2014] BrainNet, which linked three

to the coarsest possible read and write.[^yoo2013][^rao2014][^grau2014] BrainNet, which linked three

people in a shared task, is the largest network demonstrated and still transmitted a single bit per

people in a shared task, is the largest network demonstrated, and it still transmitted a single bit

sender per turn.[^jiang2019] Multi-animal "brainets" have shown that several brains can be pooled

per sender per turn.[^jiang2019] Multi-animal "brainets" have shown that several brains can be

to control one output, which is a different and somewhat less surprising result: it is population

pooled to control one output. That is a different and somewhat less surprising result, amounting to

decoding across skulls.[^ramakrishnan2015]

population decoding across skulls.[^ramakrishnan2015]

> [!caution] Why "telepathy" is the wrong word

> None of these systems transmits meaning. The bit is meaningful only because both participants were

> None of these systems transmits meaning. The bit means something only because both participants

> told in advance what it would signify. A system with the same architecture and a light bulb

> 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

> in place 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.

> percept. They do not show that brains can share content.

## Why the bandwidth is so low

**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 [[ecog-interfaces]] work show, but nobody has

substantially, as results from [[utah-array]] and [[ecog-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

cubic centimetre of cortex indiscriminately. Focused ultrasound is better localized but still

addresses large populations; intracortical microstimulation is finer but still activates hundreds

addresses large populations, and intracortical microstimulation is finer but still activates

of cells in a pattern nothing like natural activity. The read-out fidelity of current

hundreds of cells in a pattern nothing like natural activity. Current neurotechnology reads far more

neurotechnology far exceeds its write-in fidelity, which is also why sensory feedback in

faithfully than it writes, which is also why sensory feedback in [[neuroprosthetics]] lags motor

[[neuroprosthetics]] lags motor decoding. Precise write-in through optical control is possible in

decoding. Precise optical write-in is possible in animals through [[optogenetics]], but it requires

animals via [[optogenetics]], but requires genetically modifying the target neurons.

genetically modifying the target neurons.

**No common code.** Even with perfect read and write, the deeper problem is that two brains do not

**No common code.** Even with perfect reading and writing, two brains do not share a

share a representational format. The population activity encoding a particular concept in one

representational format. The population activity encoding a particular concept in one person's

person's cortex has no fixed correspondence to activity in another's; representations are shaped by

cortex has no fixed correspondence to activity in another's, because representations are shaped by

individual developmental and learning history. Delivering A's neural pattern to B would not

individual developmental and learning history. Delivering A's neural pattern to B would no more

reproduce A's experience in B, any more than installing one computer's memory contents in another

reproduce A's experience in B than copying one computer's memory contents into a machine with a

with a different architecture would run the program.

different architecture would run the program.

The plausible workaround is learning rather than translation. A receiver could, in principle, learn

The plausible workaround is learning rather than translation. A receiver could in principle learn to

to interpret an arbitrary but consistent stimulation code, in the way that users of

interpret an arbitrary but consistent stimulation code, much as users of [[sensory-augmentation]]

[[sensory-augmentation]] devices come to experience a substituted signal as a perception rather

devices come to experience a substituted signal as a perception rather than a puzzle. Nobody has

than a puzzle. That has not been attempted at any scale for brain-to-brain transmission, and it

tried this at any scale for brain-to-brain transmission, and it would make the channel a new sense

would make the channel a new sense to be learned rather than a shortcut past learning.

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

receiver intending it. Participants in the human studies consented to exactly that, but the

arrangement makes the question of agency concrete in a way that a cursor decoder does not. Related

arrangement makes the question of agency concrete in a way 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

The framing problem persists. Commercial neurotechnology has borrowed the vocabulary, and

— [[neuralink]] markets its cursor-control product under the name Telepathy — in ways that inflate

[[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

public expectations of what neural interfaces do. Typing with a decoder and sharing an experience

sharing an experience is not a matter of degree.

differ in kind, not in degree.

## Outlook

No path from the current demonstrations to meaningful brain-to-brain communication has been

No path from the current demonstrations to meaningful brain-to-brain communication has been set out

articulated in technical detail. The requirements are a high-resolution write channel that does not

in technical detail. It would need a high-resolution write channel that does not exist, a solution

exist, a solution to the shared-code problem that no one has proposed beyond "the receiver learns",

to the shared-code problem that no one has proposed beyond "the receiver learns", and an ethical

and an ethical case for implanting healthy people that nothing currently justifies.

case for implanting healthy people that nothing currently justifies.

The idea nonetheless persists because it sits at the centre of several other speculative

The idea persists anyway because several other speculative programmes depend on it. Proposals for

programmes: proposals for [[human-ai-merger]] often assume a rich bidirectional neural channel,

[[human-ai-merger]] often assume a rich bidirectional neural channel; scenarios built on

scenarios built on [[whole-brain-emulation]] assume that a mind's content can be extracted and

[[whole-brain-emulation]] assume that a mind's content can be extracted and moved; and arguments

moved, and arguments about merged or networked minds in the [[mind-uploading]] literature take

about merged or networked minds in the [[mind-uploading]] literature take shared access between

shared access between substrates for granted. Those arguments generally inherit the assumption that transmitting

substrates for granted. Those arguments generally inherit the assumption that transmitting neural

neural activity transmits its content. The brain-to-brain experiments are the closest thing to a

activity transmits its content. The brain-to-brain experiments are the closest thing to a direct

direct test of that assumption, and what they show is a single bit crossing a gap that everything

test of that assumption. What they show is a single bit crossing the gap, interpretable only because

else about the two brains has to be arranged in advance to interpret.

everything else about the two brains was arranged in advance.

## See also