In 2019, researchers at the University of Washington published a paper in Scientific Reports describing BrainNet—a system that allowed three people, seated in separate rooms with no ability to see, hear, or talk to each other, to collaboratively play a Tetris-like game using only their brain signals. Two “senders” could see the game board and decided whether a falling block needed to be rotated. They communicated their decisions to a “receiver” who couldn’t see the board but controlled the game. No words. No gestures. No screens shared between them. The senders’ decisions were extracted via EEG, transmitted over the internet, and delivered to the receiver’s visual cortex via transcranial magnetic stimulation, where they appeared as flashes of light—phosphenes—that the receiver interpreted as instructions. Five groups of three people tested the system and achieved 81 percent accuracy.
That’s the headline. Here’s the fine print: the information transmitted was binary. Yes or no. Rotate or don’t rotate. One bit of data per transmission cycle. The senders communicated their decisions by staring at lights flashing at different frequencies—15 hertz for one answer, 17 hertz for the other—which entrained their brain’s electrical output at the corresponding frequency, readable by EEG. The receiver experienced either a flash of light (rotate) or no flash (don’t rotate). The “brain-to-brain communication” was, functionally, a very elaborate way to send the equivalent of one binary digit from one head to another. IEEE Spectrum described an earlier version of this approach as “telepathic Morse code.”
This is what brain-to-brain communication actually looks like in 2026: technically real, scientifically genuine, and approximately as far from telepathy as a tin-can telephone is from a 5G network.
What exists
The field has produced a series of legitimate demonstrations, each constrained by the same fundamental bottleneck: you can get information out of a brain with reasonable resolution using EEG or implanted electrodes, but you can deliver information into a brain noninvasively only through crude channels—magnetic pulses that trigger phosphenes (perceived flashes of light) or vague sensations. The input side is the constraint. Reading a brain is hard. Writing to a brain is harder by orders of magnitude.
The 2014 Starlab experiment was the first reported human brain-to-brain transmission. A sender in India imagined moving his hands or feet to encode binary data through EEG. The signal was emailed to France, where a TMS device delivered pulses to a blindfolded receiver’s visual cortex, producing phosphenes. The receiver reported the flashes verbally, and the team decoded the message. The transmitted words: “hola” and “ciao.” The transmission rate was approximately two bits per minute. The entire process took over an hour.
BrainNet in 2019 scaled the architecture to three people and demonstrated something genuinely interesting beyond the binary channel: when the researchers injected noise into one sender’s signal, the receiver learned to preferentially weight the more reliable sender—a trust calibration process that happened entirely through brain-to-brain signals without any conscious strategy. The receiver’s brain was doing signal integration across two noisy sources, the same computation that underlies sensory integration in normal perception.
Invasive brain-computer interfaces—Neuralink, Synchron, Blackrock Neurotech—are advancing rapidly on the reading side. Neuralink implanted its first human chip in January 2024 under its PRIME study, enabling a paralyzed patient to type and control a cursor through thought alone. Synchron’s Stentrode sits inside a blood vessel near the brain, avoiding open surgery. The PRIME study has a primary completion date of 2026 and full study completion projected for 2031. These systems are brain-to-computer interfaces, not brain-to-brain—they translate neural signals into digital commands for external devices. But they represent the reading infrastructure that any brain-to-brain system would eventually need.
On the AI-assisted decoding side, researchers at the University of Texas in 2023 used fMRI scans and large language models to decode continuous thought into coherent text—not single words or binary choices but streams of semantic content, capturing the gist of what a person was thinking about during a story or imagined narrative. Meta has developed noninvasive brain-scanning systems paired with AI models that can decode silently spoken words from brain activity. These aren’t brain-to-brain systems, but they’re solving the bandwidth problem on the reading end: extracting richer, more nuanced information from neural signals than EEG-based approaches can achieve.
What doesn’t exist
Telepathy—the transmission of complex thoughts, images, emotions, or experiences from one mind to another—is not close. The demonstrations that exist transmit binary decisions through artificial sensory channels. The receiver doesn’t “hear” the sender’s thought. The receiver sees a flash of light and interprets it according to a pre-agreed code. The brain-to-brain interface is a translation chain: thought → EEG signal → digital encoding → internet transmission → TMS pulse → phosphene → interpretation. At every link in that chain, information is lost. What arrives in the receiver’s brain is not a thought. It’s a stimulus—a magnetically induced visual artifact that carries one bit of information about the sender’s decision.
The gap between this and actual telepathy is not a gap that incremental engineering will close, because the limiting factor isn’t the technology between the brains. It’s the fundamental problem of neural encoding: we don’t know, for any given thought, which specific neural firing patterns represent it, how those patterns vary between individuals, or how to induce a specific firing pattern in a target brain that would be experienced as the same thought. Brains aren’t standardized hardware. The neural code for “rotate the block” in one person’s motor cortex is not the same pattern in another person’s motor cortex. Translating one person’s neural representation into a stimulus that would produce the same internal experience in another person requires a mapping between two unique neural architectures—a problem neuroscience hasn’t solved and isn’t close to solving.
What BCI companies are building toward is not telepathy but increasingly high-bandwidth brain-to-computer interfaces that could, in principle, be linked: Brain A → computer → Brain B. Neuralink’s implant reads neural signals at thousands of channels. Future implants will read more. AI decoding systems are getting better at extracting semantic content from neural data. But the write side—delivering complex, precise, meaningful information directly into neural tissue in a way that the receiving brain interprets as a coherent experience—remains the unsolved problem. TMS can trigger phosphenes and crude sensory impressions. It cannot implant a sentence, an image, an emotion, or a memory.
The timeline problem
Coverage of brain-to-brain communication tends to imply a trajectory: binary transmission today, sentences tomorrow, telepathy eventually. The trajectory is real in the same way that the Wright Brothers’ 12-second flight in 1903 implied commercial aviation—the physics supports the possibility, but the engineering required to get from demonstration to deployment is measured in decades, not years, and the technical obstacles on the write side are qualitatively different from the obstacles on the read side.
Reading a brain is an information extraction problem: the neural signals are there, and the challenge is building sensors sensitive enough and algorithms smart enough to decode them. This problem is yielding to better hardware and better AI. Writing to a brain is an information implantation problem: you need to induce specific patterns of activity in specific neural populations at specific times, through skull and tissue, without disrupting the brain’s existing activity. Noninvasive methods (TMS, focused ultrasound, transcranial electrical stimulation) affect large regions of cortex with limited spatial precision. Invasive methods (optogenetics, direct electrical stimulation) can target individual neurons but require surgery, gene therapy, or implanted hardware.
The honest assessment in 2026: brain-to-computer interfaces are advancing on a trajectory that will produce clinically meaningful products for paralysis, communication disorders, and sensory prosthetics within the current decade. Brain-to-brain communication, in the sense of transmitting complex mental content between two people, requires solving the neural write problem at a resolution and precision that current technology can’t achieve and that current neuroscience can’t specify. The demonstrations are real. The extrapolation to telepathy is premature by a margin that is difficult to estimate because the bottleneck isn’t engineering velocity. It’s a scientific knowledge gap about how brains encode experience—a gap that better instruments may close but that no existing roadmap guarantees.
Neuralink named its first consumer product “Telepathy.” The name is aspirational in the way that calling the first automobile a “teleporter” would have been aspirational. The product lets a paralyzed person control a cursor with their thoughts. That’s extraordinary and useful. It’s not telepathy. The distance between the two is the distance between reading a book and writing one—and in neuroscience, we’re still learning to read.
We cover brain-to-brain communication alongside spinal cord stimulation, retinal implants, and the full landscape of neural interface technology across our Neuroprosthetics course—including why the hardest problem in connecting two brains isn’t getting the signal out. It’s getting the signal in.