A building has no idea what you mean. It knows only what you press. For a century, every light switch, door opener and lift call has turned an intention into a hand movement and then into a closed contact. A brain-computer interface (BCI) skips the hand. It reads the intention as an electrical pattern in the cortex and sends a command. The idea is about a hundred years old. What changed in this decade is that it left the lab bench and reached real people, which means it now enters procurement, consent forms and control buses.

The essay below covers what a BCI is, why it works, who built it, and what a Swiss office should specify before the first client asks for one. I write from the late 2070s, and I watch the governance layer. The interesting failures in this field were never in the electrodes.

What it is

What it is: a BCI is a closed-loop control system in which one of the controllers is a person. The PAZ concept panel on BCI engineering describes the chain, and it is the same whether the sensor is a scalp cap or a 1,024-electrode implant: acquisition, filtering, feature extraction, decoding, device command, feedback. The last step is what makes it work. The user sees the cursor move, corrects it, and over time the brain and the decoder adapt to each other. Engineers call this co-adaptation. Nothing here is telepathy. It is pattern recognition on voltages, with a human in the loop who keeps learning.

Why it works

Why it works: motor intent leaves a measurable trace before any muscle moves. When you imagine closing your left fist, the mu (8–12 Hz) and beta (13–30 Hz) rhythms over the right motor cortex weaken. Neurons in the motor cortex also fire in population patterns that encode direction and speed. A decoder learns how those patterns map to commands. Early intracortical decoders were linear: population vectors first, then Kalman filters that estimate cursor velocity from firing rates. From 2021 to 2023, recurrent neural networks combined with language models changed the scale of what was possible. At Stanford, Frank Willett, Krishna Shenoy and Jaimie Henderson decoded imagined handwriting at 90 characters per minute (Nature, 2021) and attempted speech at 62 words per minute (Nature, 2023). Edward Chang’s team at UCSF reached 78 words per minute with a digital avatar (Metzger et al., Nature, 2023). As Smithsonian reported, the Chang Lab has since shown an implant that lets a person with paralysis speak and gesture through the avatar at the same time. Most earlier systems did only one or the other.

The physics comes with a trade-off you cannot get around. Skull and scalp blur the signal like a thick diffuser. Scalp EEG is cheap and safe, but it sees only summed microvolt activity from cortical patches several centimetres wide. Each step closer to the neurons improves the signal: epidural ECoG grids such as WIMAGINE (64 electrodes), endovascular arrays such as Synchron’s Stentrode (16 electrodes, placed through a vein), and microelectrode arrays such as the Utah array (about 96 channels) or Neuralink’s N1 (1,024 electrodes on 64 threads). Every electrode you move closer to the neuron buys bandwidth and costs a surgery, years of biocompatibility work, and a signal that may degrade over time.

Accuracy figures can mislead, so use Jonathan Wolpaw’s information transfer rate as the yardstick. For a two-class motor-imagery BCI, 80 % accuracy carries only 0.28 bits per decision. At 95 % it carries 0.71 bits. A push button gives almost a full bit in under a second. An 80 % BCI is therefore not almost a button. It is a different kind of input and needs a different kind of design.

←TODAY: in 2026, implanted BCIs serve a few dozen people, and building controls still assume a hand on a KNX push button.
→3012: intent becomes one routine channel on the building bus, next to voice, gaze and touch, and no channel is allowed to be the only one.
Fulcrum: the rule that keeps a BCI humane is the same rule that keeps it governable: redundancy of input means nobody is locked in, technically or contractually.

Origins

Origins: the story starts in Jena. In July 1924 the psychiatrist Hans Berger recorded the first human electroencephalogram. He kept checking his results for five years before publishing them in 1929 as “Über das Elektrenkephalogramm des Menschen”. The alpha rhythm he described is still the first trace students see. In 1973 Jacques Vidal at UCLA gave the field its name and its question in “Toward Direct Brain-Computer Communication”: can brain signals become a channel for a computer, the way a keyboard is?

It took three decades to get from that question to the clinic. Philip Kennedy placed early implants in the late 1990s. In 2004 the BrainGate pilot trial began in John Donoghue’s group at Brown, using the microelectrode array that Richard Normann developed at the University of Utah. Matthew Nagle, paralysed from the neck down, moved a cursor, opened e-mail and closed a prosthetic hand by thinking (Hochberg et al., Nature, 2006). Credit belongs to him and to the trial participants who followed. They spent years as the instrument while the decoders learned, and the published speed records were built on that work.

The Swiss chapter is the one this readership should know by name. Grégoire Courtine at EPFL and the neurosurgeon Jocelyne Bloch at CHUV built a “digital bridge”. A wireless WIMAGINE implant from Clinatec in Grenoble reads the motor cortex and drives stimulation of the spinal cord below the injury. In 2023 it let Gert-Jan Oskam walk naturally again more than a decade after a cycling accident (Lorach et al., Nature, 2023). ONWARD Medical, an EPFL spin-off, is carrying the line toward the clinic. In Geneva, the Wyss Center team that included Ujwal Chaudhary and Jonas Zimmermann helped a man with ALS in a completely locked-in state spell out sentences (Nature Communications, 2022). Synchron reaches the motor cortex without opening the skull. Neuralink first implanted the N1 in a person in January 2024.

The engineering frontier is still moving, and much of it is European. A BioSpace release describes how INBRAIN Neuroelectronics and the MINIGRAPH European consortium are developing magnetically guided robotic implantation for ultra-thin graphene interfaces. Meanwhile, as WIRED reported, a new company, Bridge Neurotech, plans a wearable BCI based on ultrasound. Venture money is now flowing toward the no-surgery end of the curve.

The control layer nobody draws

Look at who controls the system. The patient carries the electrodes. The decoder weights, the firmware update channel, the calibration data and often the cloud account sit with the manufacturer. Neural data is the most intimate telemetry a person can produce, and the default in consumer technology has been to collect first and ask later. Chile wrote neural rights into its constitution in 2021. Colorado added neural data to its privacy law in 2024. Switzerland’s revised data protection act (nDSG, in force since September 2023) treats health data as sensitive personal data. Implanted devices fall under medical-device regulation. A building controller that receives the decoded command falls under none of these by design.

That gap is where I would point a present-day reader. From where I write, the lasting harm did not come from a decoder that misread a thought. It came from installations where an assistive input was tied to one vendor’s gateway and nobody had written down how to leave. Write the exit clause before the entry contract.

In practice

In practice: once intent is decoded, the building side is ordinary. Doors, lifts, lights and blinds already sit on KNX (EN 50090, ISO/IEC 14543-3) or BACnet (ISO 16484-5). A BCI is one more sensor on that bus, next to the push button, the voice assistant and the eye tracker. The offices that will meet it first are a hospital planner working on a CHUV-type rehabilitation brief, an accessibility consultant applying SIA 500, and an electrical planner writing the controls specification for assisted living. SIA 500 and ISO 21542 describe how people with different bodies reach controls. Neither yet describes an input that uses no hand at all. Two design problems are already known. The first is the Midas touch: how often does the door open when the user is only thinking about the door? Robert Jacob described it for gaze input in 1990. The answers are known too: a “brain switch” that first detects the intention to act (Pfurtscheller’s work), dwell or confirmation steps, and hybrid inputs. The second problem is spatial. At CHUV, the operating theatre, the rehab gym and the gait lab work as one instrument, so the distances between those rooms become part of the therapy. Synchron’s COMMAND trial placed a Stentrode at Mount Sinai in July 2022 in an interventional suite, not a neurosurgical theatre. The rooms that can host a BCI are becoming smaller and more common.

Atelier: for a Swiss Büro, adopting this is not a hardware decision. It is a specification decision, and it belongs in the same template where you already write requirements for voice and touch controls. The risk is not the implant. The risk is a proprietary gateway quietly becoming the only way one occupant can open their own front door. Monday move: add one clause to your controls specification template requiring that any non-manual input (voice, gaze, neural) reach the bus through an open KNX or BACnet gateway, that raw biosignals never leave the user’s device, and that every action stays reachable through at least one other physical path.

Hack

Hack: measure a BCI in bits per minute, not in accuracy, before anyone in the meeting says “95 % is good enough”. The function below implements Wolpaw’s information transfer rate for N choices, accuracy p, and t seconds per decision. Compare 80 % and 95 % at four seconds per trial and you get about 4.2 against 10.7 bits per minute. That is the real difference between a frustrating door and a usable one. Keep p strictly below 1, because log2(0) is undefined. In the PAZ Building System Specialist track, this is the number we put next to a push button’s roughly one bit per second.

from math import log2
def itr(n, p, t):  # bits/min; n choices, accuracy p (<1), t s per decision
    b = log2(n) + p*log2(p) + (1-p)*log2((1-p)/(n-1))
    return b * 60 / t
print(itr(2, 0.80, 4), itr(2, 0.95, 4))

Run it with the accuracy a vendor quotes you, put the result next to the 60 bits a minute of a button pressed once a second, and write both into the specification.