The Mind as a Control System

Would it be possible for a machine to hold an estimate of how a person is doing, notice that the estimate is drifting, and nudge it back the way a thermostat nudges a room toward a set temperature? The question sounds like science fiction, but a narrow version of it has already been tested in human beings. Understanding exactly what those tests did and did not show is the fastest route to understanding both the promise and the danger of treating a mind as a control system.

A patient, a signal, and a switch

In October 2021, a team at the University of California, San Francisco described a single case in Nature Medicine (Scangos et al., 2021). The patient was a 36-year-old woman with severe, treatment-resistant major depression who had not responded to multiple antidepressant combinations or to electroconvulsive therapy. Over a ten-day mapping period, the researchers recorded her brain activity from electrodes threaded through several corticolimbic regions and, at the same time, collected repeated symptom ratings. They then trained classifiers to separate high-symptom states from low-symptom states. A single feature did most of the work: the power of gamma-band activity in the amygdala identified the high-symptom state with roughly 77 percent accuracy.

The team then took the decisive step. They implanted a responsive neurostimulation device that continuously listened to that amygdala signal and delivered a brief pulse of electricity to a connected site only when the pathological pattern appeared. The stimulation was not constant. It was triggered. When the loop was closed, the authors reported, the patient’s depression improved rapidly and durably. They were also explicit about the limits: this was an n-of-1 study, one person and one device, and a single case cannot establish how many other patients would respond the same way.

Two years later, a larger study moved the idea from one patient to a small cohort. In Nature, a separate research team reported on ten people with treatment-resistant depression who received deep brain stimulation of the subcallosal cingulate (Alagapan et al., 2023). At 24 weeks, nine of the ten showed a robust clinical response and seven reached remission. More revealing than the response rate was the method applied to it. Using recordings from the implanted device, the researchers trained an explainable model to distinguish a person’s current recovery state from the momentary ups and downs of stimulation itself. The resulting biomarker was sensitive to therapeutic adjustments and tracked individual recovery rather than a single universal template. In a field where clinicians have long relied on interview scales and intuition, an objective, individualized readout of “how this patient is doing right now” is a genuine advance.

These two studies are worth holding onto because they mark a real boundary. In controlled, invasive, clinical settings, the loop from measurement to inference to intervention to re-measurement has already been closed in humans. Everything else in this article sits on one side or the other of that boundary.

What “control system” actually means

The phrase control system is engineering shorthand, and it is worth unpacking before it gets inflated. A control system has a few recurring parts. There is a setpoint or reference — the value you want to hold. There is a sensor that measures something about the world. There is an estimator that turns noisy measurements into a best guess about the true state, since what you can sense is rarely the thing you actually care about. There is a controller that compares the estimate with the setpoint and decides what to do. There is an actuator that carries out the decision. And there is a plant — the system being controlled — along with disturbances, the outside influences that push the plant away from the setpoint. The loop closes when the effect of an action is measured and fed back into the next decision.

A thermostat is the canonical example. The setpoint is the temperature you dial. The sensor is the thermometer. The actuator is the furnace. The disturbance is an open window. Crucially, even in this trivial case there is an estimator, because the thermometer reads the air near the wall, not the whole room.

Two refinements matter for the brain. The first is the difference between open-loop and closed-loop control. Open-loop stimulation delivers a fixed dose on a fixed schedule and never checks the result. Closed-loop stimulation measures the state and acts only when needed, exactly as the UCSF device acted only when the amygdala signal crossed a threshold. Closed-loop systems can be gentler, because they avoid intervening when the system is already in an acceptable range, and they can reduce side effects and energy use by staying quiet much of the time.

The second refinement is model-based control, sometimes called model predictive control. Instead of reacting only to the present error, the controller uses an internal model of the plant to forecast what will happen under different actions and chooses the action that looks best over a short horizon. This is where machine learning has changed the picture. A system that learns a model of a specific person’s dynamics can, in principle, anticipate rather than merely react.

None of this vocabulary is sinister. It is the same vocabulary used for cruise control, insulin pumps, and spacecraft. The question is never whether the vocabulary applies. It is what, specifically, is being controlled, toward what objective, and by whom.

Why a brain is not a furnace

The thermostat analogy breaks down almost immediately, and the ways it breaks down are the real subject.

A furnace is roughly linear: more fuel produces more heat, predictably and in proportion. The brain is not. Neural circuits are nonlinear, densely interconnected, and dependent on their current state. A stimulation parameter that calms a circuit in one condition can excite it in another. This is not a technicality. A 2022 review in Nature Reviews Neuroscience catalogued the ways the effect of the same neural stimulation depends on the state the brain is already in (Bradley et al., 2022). The practical consequence is that “the right dose” is not a constant. It is a function of context.

The brain is also plastic. It adapts to its inputs. A controller that assumes the plant is fixed will drift out of calibration: the system it learned to steer is no longer quite the system it is steering. This is one reason closed-loop stimulation can outperform open-loop stimulation — by stimulating less, it may slow the adaptation that erodes the response over time.

Most important, the thing we care about is not directly measurable. “Depression,” “calm,” and “focus” are not single physical quantities. What a sensor can deliver are proxies: amygdala gamma power, heart-rate variability, pupil diameter, reaction time, self-reported mood. A proxy is not the thing itself, and treating it as if it were is the most common and most dangerous error in the entire field. Amygdala gamma is not sadness. Heart-rate variability is not composure. A model that estimates the proxy well may still be wrong about the person.

That gap — between what is measured and what is meant — is where the ethics of control systems actually live.

Measuring an inner state

The honest starting point is that inner states can be estimated from brain activity better than most people would guess, but far less precisely than the headlines imply. In 2018, a team reported in Nature Biotechnology that it could decode moment-to-moment mood variation from multi-site intracranial recordings taken from people who already had electrodes implanted for clinical reasons (Sani et al., 2018). The finding is real and it is striking: mood, one of the most subjective experiences a person has, left a decodable trace in coordinated activity across several brain regions rather than in one spot.

But note the conditions under which that was possible. The recordings were invasive, taken from a small number of participants, and interpreted with custom models built for those individuals. The result tells us that mood has a neural correlate that is, in principle, observable. It does not tell us that a consumer device can read a person’s feelings from the outside, and the distance between those two statements is enormous.

For most practical systems, the inputs will be indirect — wearables, cameras, voice, behavior — and the estimate will carry real uncertainty. A responsible design keeps that uncertainty visible rather than collapsing it into a confident label. Knowing that the estimate might be wrong, and roughly how wrong, is what allows a system (or a clinician, or the person) to decide whether to act at all.

One patient at a time

If there is a single lesson running through the clinical literature on closed-loop neuromodulation, it is that individual variability is the central problem and personalization is the only serious answer.

That theme is clearest in a 2025 randomized trial reported in Nature Communications, which tested deep brain stimulation targeting the bed nucleus of the stria terminalis and nucleus accumbens in 26 patients with refractory depression (Wang et al., 2025). In the open-label phase, half the patients responded and about a third reached remission. But the study’s more consequential contribution was predictive. By combining intracranial recordings, machine learning, and imaging, the team found that specific physiological features — lower theta activity in one target region and a particular pattern of prefrontal connectivity — predicted who would do better over the following year. In other words, the study did not simply ask whether stimulation worked. It began to ask, for a given person, whether this particular intervention was likely to work, before committing to years of treatment.

This is what a mind-as-control-system looks like in the clinic: not one universal setting applied to everyone, but an individualized model, an individualized setpoint, and an individualized trigger. The same logic underlies the adaptive deep brain stimulation already approved for movement disorders. A multicentre, nonrandomized 2025 Parkinson’s trial in JAMA Neurology found that adaptive stimulation, which adjusts to the brain’s own signals, reduced the total electrical energy delivered while improving the amount of symptom-controlled time (ADAPT-PD, 2025). That is the closed-loop promise in its most honest form: less intervention, better timed, tuned to the person.

From the clinic to daily life

Everything above takes place inside a clinic, with implanted hardware, consenting patients, and clinicians watching. The speculative leap — the one that makes this series necessary — is from treating disorders to tuning ordinary inner life, and from invasive hardware to something a person might simply wear or use.

The scientific groundwork for non-invasive deep control is more advanced than skeptics assume and less advanced than enthusiasts claim. Transcranial ultrasound stimulation can now reach deep structures through an intact skull. A 2025 study in Nature Communications took 26 healthy adults and used focused ultrasound aimed at the nucleus accumbens, a reward-related region, while participants performed a learning task in a scanner (Nature Communications, 2025). The stimulation changed reward-related brain responses and behavior — how quickly participants updated their choices after a reward — and importantly, stimulating a control region produced different effects, which argues for specificity rather than a generic jolt. Temporal interference, which uses interacting electric fields to reach depth without surgery, has followed a similar path from animal work and modelling into early human studies (Vassiliadis et al., 2026).

Read these results for what they are. Researchers changed a measurable, reward-related response in healthy volunteers under controlled conditions. They did not install a mood dial, and no consumer device can currently tune a person’s emotional state from the outside. Non-invasive methods still face hard problems: the skull blurs and attenuates what reaches the brain, dosing is capped by safety limits, and anatomy differs enough between people that targeting precision varies. The demonstrated science and the popular scenario are separated by years of engineering and a great deal of unresolved safety work. A responsible account keeps them apart.

The best case

The positive case for controlled neural systems is not thin, and it would be a mistake to let the risks crowd it out. Communication decoders that restore a voice to someone who has lost speech are a straightforward good. Closed-loop stimulation that fires only when a patient’s own brain signals indicate an oncoming episode is more precise, gentler, and more respectful of the person’s variable condition than a constant dose. As the Parkinson’s trial suggests, adaptivity can mean less total energy delivered, which may help conserve battery energy; side effects and battery longevity still require direct assessment.

There is also a subtler benefit. A closed-loop system turns a person’s own biology into the reference signal. Instead of imposing an outsider’s average, it learns the individual’s patterns and responds to them. Done well, this is the opposite of coercion: it is a treatment that adapts to the patient rather than forcing the patient to adapt to the treatment.

The failure mode

The dangers begin the moment we ask a deceptively simple question: what is the setpoint, and who gets to set it?

A controller needs an objective, and the objective is a choice. A system optimized for a patient’s stated values is not the same as one optimized for productivity, engagement, sales, or institutional compliance — even when the underlying hardware is identical. That is not a hypothetical concern. A system told to maximize productivity may discover that the cheapest way to hit the target is to suppress fatigue, and a system told to maximize engagement may find that mild anxiety and intermittent reward keep a user returning. Intelligence does not remove the need to choose a wise objective. It makes the objective more consequential, because the system will pursue it more effectively than any fixed tool.

Consent is the second fault line. It can be weak when the person has no realistic alternative — an employee, a student, a patient, a customer. It becomes genuinely unstable when the intervention changes the very preference on which consent was based. If a device alters how a person feels about using the device, then “the user is happy with it” stops being a clean test of whether the user would want it. Meaningful choice, in this domain, has to include the ability to inspect what the system is doing, to override it, and to disconnect without penalty.

The third fault line is identity. If a device changes the internal state from which a decision emerges, whose decision is it? A 2025 scoping review in npj Digital Medicine examined 66 clinical studies of closed-loop neurotechnology and found that only one included a dedicated ethical assessment; where ethical language appeared, it was usually procedural — a reference to review-board approval — rather than a substantive examination of autonomy, identity, or privacy (Haag et al., 2025). The technology is moving faster than the reflection on it. That gap is itself a finding.

Assistance and control

It is worth stating the distinction this series keeps returning to in concrete terms, because “control” is used loosely to mean two very different things.

A system assists when it widens a person’s range of voluntary action: it restores a function, removes an obstacle, or offers an option the person can take or leave. A system controls when it narrows that range, even if the narrowing feels convenient. The same device can do either, depending on where the objective comes from and whether the person can inspect and veto it.

This gives a test that survives contact with real technology. Ask who chose the setpoint, whether the person can see and change it, whether the effect is reversible, and whether the person becomes more capable over time or merely more dependent on the system. A treatment that helps someone reclaim their own goals passes. A system that quietly substitutes someone else’s goals — a company’s, a state’s, an advertiser’s — fails, however pleasant it feels from the inside.

Where this leaves us

The narrow engineering question is largely answered. Given clean signals, careful personalization, and clinical oversight, an AI can estimate a person’s state, decide when intervention is warranted, act, and learn from the result. That has been demonstrated in humans, and the results are not trivial. They are, in several cases, the difference between treatment-resistant illness and remission.

The larger question is not technical at all. It is who defines the objective and whether the person retains real authority over it. The technology of the loop is arriving faster than the governance of the loop. Between those two clocks runs the entire difference between a mind that is supported and a mind that is steered — between an instrument that serves a person’s own aims and a sovereign layer that decides, quietly and individually, what those aims should be.