For as long as there have been people, the inside of the skull has been the one room nobody else could enter. A technology that reads from it has to earn the right to exist.

The near versions do not require reading a mind. They require a system that can tell a deliberate act from everything else a nervous system does, and that treats the difference as sacred. That system is the intent layer: sensing, interpretation and learning, behind a gate only the user can open.

Who holds the interface

BCI began as a clinical technology. It is becoming a productivity technology. It will become a data technology once neural recordings count as training material, and then the ordinary channel between people and machine intelligence. Its strategic weight rises at each step.

States have noticed. One major power names BCI among its strategic future industries. Another leads in startups and frontier AI. A third is writing the rules for mental privacy that others will have to meet. Defence agencies have funded the field for decades. The AI decade was a contest over chips. The next will include the sensors that record the nervous system, and the question of who may collect, model and act on what they record.

We do not think this should be a technology a few countries own. The people with most to gain from a cheap interface live everywhere and speak every language. Open standards and devices priced for ordinary incomes are how it reaches them.

What can go wrong

• A recording says more later than it does now. A trace that shows only drowsiness to this year’s models may show mood or health to the models of five years on. Consent given today covers inferences nobody can yet list.
• Manipulation. A system that can sense a reaction can learn to produce one.
• Drift of the self. Did I choose that, or was I completed?
• Intrusion. Everything connected can be attacked, and this is the most intimate attack surface ever shipped.
• Surveillance. Monitoring attention at work or school, profiling dissent, interrogation. Someone will propose each.
• Inequality. An interface to machine intelligence owned by a few and priced for a few.

The state of practice is not reassuring. A review of thirty consumer neurotechnology companies found that nearly all reserved broad rights over the neural data they collect.⁶²

How this goes right

Good outcomes have to be designed, and the strongest safeguards are technical before they are legal. A law can be repealed. An architecture has to be rebuilt.

• Process on the device. Send intent, not waveforms.
• Keep neural data away from advertising. Always.
• Make consent expire. Next year’s models will read what this year’s cannot.
• Give neural data its own legal category. No resale, no hidden inference, no access for employers or insurers.
• Hold writing to a higher bar than reading.
• Agree the red lines early. No neural interrogation, no compulsory monitoring, no covert stimulation.

The work has started. Neuroscientists set out the case for mental privacy in 2017.³¹ One country has written protection of brain activity into its constitution.³³ In November 2025 the member states of UNESCO adopted the first global standard.³² For once, rules may arrive before the scandal.

There is a failure on the other side too. Regulate the field to a standstill and the result is not safety. It is a frozen technology in the years it matters most.

What Setu holds itself to. Sensing is off unless the user turns it on. Raw signal stays on the device wherever the engineering allows. We decode what was deliberately offered and nothing else. Every output can be seen, corrected and undone before it acts. Neural data will not be used for advertising. Nobody should have to wear our product as a condition of a job or a classroom.

The case against us

The keyboard has survived fifty years of predictions. Hands and voices are superb.

They are, and they did not evolve to direct machines a thousand times faster than they can type. BCI starts where hands and voice fail and grows from there.

These results come from single patients in heroic laboratory conditions.

Correct. One participant is an existence proof. It shows the information is there. Turning it into something a stranger can wear is the rest of this essay.

The skull is a filter. No dataset repairs physics.

The skull sets a ceiling somewhere. Nobody has found it. Every EEG curve measured so far is still rising. Speech recognition was not forbidden by physics either. It needed a hundred thousand hours.

Effort spent outside the skull is taken from implants, which have the higher ceiling.

For many years only a device without surgery can reach millions. The methods, the manufacturing skill and the people it produces carry across.

The brain is not the internet. There is no objective as clean as predicting the next word.

The objective is prediction across signal and context: given a recording and what the person saw, heard and did, predict the missing part. AI has scaled on messier objectives.

Consumer EEG headsets have been tried. The category is a graveyard.

The earlier wave built products before there was a stack beneath them: no general models, no buyer for the data, no agents that made a new input valuable. Most technologies fail until the missing pieces exist.

It is too private ever to be collected at scale.

That is why the safeguards have to be in the first device. A version that treats neural data the way advertising treats browsing history deserves to fail.

The channel could strengthen the machine more than the person.

It could. Nothing about the technology favours the human automatically. That has to be designed.

What would prove us wrong

Our case rests on three assumptions. EEG carries enough information. Decoding keeps improving as data grows to a hundred thousand hours and to a million. And models trained across enough people cut each new person’s calibration from hours to minutes.

Anyone is free to disagree, and the disagreement is testable. If any of the three fails, the strong version of this thesis is wrong. We do not expect that.

The condition

The cheaper and more widely worn a sensor becomes, the more people it helps and the more intimate data exists in the world. Wider access and deeper exposure arrive together. Good design decides that tension in the user’s favour, every time, by default.

Beneath the policy sits a choice no regulator can make for us: what we want to keep about being human. Agency. Authorship. The privacy of an inner life.

This future will not be good by default. It has to be built on purpose, and not by some later generation. If you build, fund, regulate or simply care how people will live beside intelligent machines, find where you can push. We would like to hear from you.

If it is built, it will be judged by two silences. The first is the one it ends: the silence of a person who has something to say and no way to say it. The second is the one it must keep: the silence of everything a person chose not to say.

An interface that ends the first and guards the second would be a new layer of computing. One that ended both would be something else, and we will not build it.