Writing to the Brain: The Harder Half of the Problem

Imagine an instrument technician who is very good at listening. Give her a live cable and she can tell you whether it carries a signal, roughly what shape the signal has, and even, with practice, what it probably means. Now ask her to send a signal the other way — to inject a specific, intended waveform into a specific wire, in a specific circuit, so that the equipment at the far end does exactly one thing and nothing else. Listening was hard. Writing is harder, for reasons that have nothing to do with how clever the technician is.

Neurotechnology works the same way. Reading a neural state and writing one are treated in public conversation as two directions of a single capability, as if a technology that can detect a pattern could obviously also impose one. The asymmetry between them is the subject of this article, and it is not a small technicality. Detection can proceed passively, by observing what is already there. Writing requires a causally effective intervention into a system that is nonlinear, distributed, state-dependent, and constantly adapting. The same pulse of energy that nudges a circuit one way in one person, or one moment, can nudge it the other way in another. That is the core difficulty, and it organizes everything below.

Two different problems wearing one hat

The first thing to separate is measurement from control. A detector is judged by whether its estimate tracks the true state across a range of conditions. A writer is judged by whether an action reliably produces a target effect — and the standard is much stricter, because a control system must work as a cause, not merely as a correlate.

Suppose a mood decoder correctly predicts that a person’s self-reported mood is rising. That is a correlation between a signal and a label. Suppose further that the same signal is fed into a stimulator trained to keep mood steady. Now the system must know not just that the signal accompanies the mood but how a pulse of current will change it, in which direction, by how much, for how long, and with what side effects. The decoder’s accuracy tells you almost nothing about whether the stimulator will work. A model can read a system perfectly and still be unable to steer it.

This is why progress on decoding does not automatically translate into progress on writing, and why the two deserve separate articles. The rest of this piece is about what writing a neural state actually requires, and why each requirement is harder than it first appears.

Why stimulation is not a scalpel

The intuitive picture of neural stimulation is a scalpel: aim energy at a target, activate that target, leave the rest alone. In 2009, a team testing this assumption directly published a result that should have reset the field’s expectations (Histed, Bonin & Reid, 2009). Using two-photon calcium imaging to watch which neurons actually fired in response to electrical microstimulation of cortex, they found that stimulation activated sparse, distributed populations of neurons up to a millimeter away from the electrode — not a tidy sphere of tissue but a scattered cloud. Moving the electrode by as little as thirty micrometers could completely change which cells responded.

The mechanism helps explain the surprise. Electrical current delivered through an electrode does not preferentially activate the cell bodies nearest to it. It tends to activate axons — the long output fibers — which are often passing through the region rather than belonging to it. The result is that stimulation recruits a set of neurons defined by which fibers happen to run near the electrode tip, not by which cells the experimenter meant to target. Two adjacent placements can therefore produce quite different effects, and the effect of a pulse is a property of the local wiring as much as of the intended target.

This single finding reframes the whole problem. Writing to the brain is not the act of pressing a button labeled “this circuit.” It is the act of perturbing a dense, interconnected network at one small point and then living with wherever the perturbation propagates. That is a fundamentally harder design problem than reading a signal, and it is why the phrase “brain stimulation” hides more than it reveals.

The brain is not a grid of switches

The distributed-activation result points to a deeper issue: neural circuits are not independent modules arranged in a neat array. They are densely interconnected, and a target almost never participates in one function alone.

A 2022 review in Nature Reviews Neuroscience catalogued how the effect of the same neural stimulation depends on the state the brain is already in (Bradley et al., 2022). The same input delivered when a circuit is quiet and when it is busy can produce opposite results. This state-dependence is not an edge case; it is a general property. It means the concept of a fixed dose — the amount of stimulation that “works” — is not even well defined. The right dose is a function of the current context, which changes from second to second and from person to person.

The distributed nature of function compounds the problem. A decoder can observe coordinated activity across many regions and still produce a useful estimate, because observation tolerates distributed representation: it just needs to listen to all the relevant sites at once. A writer cannot so easily act on all of them. To change a distributed state, you would have to reach many sites simultaneously with the right relative timing, which is precisely what current stimulation methods cannot do. This is one reason the read side has outpaced the write side: the reading problem can be solved by listening everywhere, while the writing problem requires acting everywhere at once.

Dose, frequency, and target: the fragile trio

The proof that writing is fragile comes from experiments in which careful, well-intentioned interventions produce opposite results depending on how they are tuned. A vivid example is a 2025 study of non-invasive temporal interference stimulation, a technique that aims to reach deep structures without surgery, applied to figure-memory encoding in 70 healthy participants (Missey et al., 2025). The results read like a warning label. A higher-frequency envelope of 130 Hz aimed at the hippocampi and temporal cortices significantly impaired recall. A lower-frequency 5 Hz envelope aimed at the hippocampi alone significantly enhanced recall. Several other target combinations produced no measurable effect at all.

Read that again as a designer. The same technique, in the same study, moved memory in two opposite directions depending on a frequency parameter, and did nothing for a range of other settings. This is not a story about a machine that turns memory up and down. It is a story about how sharply the outcome depends on choices of target and dose that were not fully known in advance. A system that produces enhancement here and impairment there, with no effect elsewhere, is not yet a device. It is a research instrument mapping out a parameter space that is only partly understood.

The invasive literature reinforces the point from a different direction. In 2025, a team combined single-neuron recordings with focal electrical stimulation in the human hippocampus and found that stimulation during a working-memory task disrupted the selective, persistent neural activity that represented the memoranda, reducing accuracy and slowing responses (Daume et al., 2025). At the population level, stimulation pushed neural trajectories away from the stable attractor states associated with accurate memory. The intervention was informative and cleanly measured — but its direction was disruption, not enhancement. This is what direct stimulation most reliably does: it interferes with the very signal you are trying to support. A writer who wants to improve a state faces the problem that the blunt tool tends to blunt the function.

That pattern is not confined to memory. In a small study of five patients undergoing surgical evaluation with electrodes already in place, stimulation of the medial temporal lobe impaired recall most when it was applied during the delay between learning and retrieval, as though the intervention erased the trace it was meant to protect (Merkow et al., 2017). The most consistent finding across human stimulation studies is disruption. Enhancement is the harder and less frequent result, and it appears only under narrower, better-understood conditions.

Reaching deep without opening the skull

The appeal of non-invasive deep-brain stimulation is obvious, and the technical story is genuinely clever. Temporal interference stimulation delivers two high-frequency alternating currents through electrodes on the scalp. Each current alone is too fast for neurons to follow, so neither stimulates tissue by itself. Where the two fields overlap, they create a slower amplitude-modulated envelope — the difference frequency — and it is this envelope that neurons can respond to. By steering the fields, researchers aim to place the envelope’s maximum at a deep target while keeping it weak elsewhere.

The approach has produced real human results. In 2023, researchers reported that temporal interference stimulation of the human hippocampus could focally modulate hippocampal activity in 20 healthy participants while they performed a memory task, and that longer stimulation in a separate behavioral experiment improved the accuracy of episodic memories. Cadaver measurements supported the targeting claim: the modulated envelope was roughly 75 percent stronger in the hippocampus than in the overlying cortex (Violante et al., 2023). The result matters because it showed that a deep, non-invasive target is not pure fiction.

It also showed where the field stands honestly. By 2025, a systematic review found that of 127 publications screened, only 18 human studies met its criteria, and that the evidence was “preliminary and predominantly focused on cortical rather than deep targets” (Mansourinezhad et al., 2025). A separate systematic review of human temporal interference applications reached a similar verdict: early-phase studies suggest deep targets can be engaged, but well-controlled Phase 2 trials are still needed to establish real clinical benefit (Demchenko et al., 2025). Dosing conventions vary from study to study, some findings on brain activation are not consistent with each other, and a fixed electrode montage can fail across individuals whose anatomy differs. The route to deep structures exists. The map is still being drawn.

Building a set of tools

None of this means writing to the brain is impossible. It means no single tool covers the space, so the field has grown a range, each trading one limitation for another.

Deep brain stimulation reaches subcortical targets with precision by placing electrodes directly in them, at the cost of surgery and its risks. Transcranial magnetic stimulation induces currents in cortex from outside the skull, with good temporal control but shallow reach. Direct electrical stimulation during neurosurgery gives unmatched spatial precision and access to otherwise unreachable sites, but only during a clinical procedure. Focused ultrasound can steer energy to depth without surgery, though the skull distorts and attenuates what passes through it. Temporal interference offers non-invasive depth in principle, with the precision questions above. Sensory entrainment and peripheral nerve stimulation work indirectly, through the body’s own pathways.

The pattern is that every method that improves reach makes control harder, and every method that improves precision pays in invasiveness or in limited access to the brain. That is not a temporary state of affairs. It is a consequence of what the skull does to energy and of how neurons respond to it.

Why the asymmetry is fundamental

It is worth stating the asymmetry plainly, because it explains why this is the harder half of the problem and not merely the next step.

Detection tolerates ambiguity. A decoder can aggregate many noisy signals, weight them, and output a probability. Distributed and state-dependent coding are helpful to a decoder: more sites mean more information to combine, and a state-dependent signal can be disentangled with enough data. A detector is allowed to be uncertain, as long as it reports how uncertain it is.

Writing does not have that luxury. An intervention must commit to a specific action at a specific place and time, and the effect of that action depends on the same distributed, state-dependent, plastic system that made detection hard. Causality is not symmetrically recoverable from correlation. Knowing that activity in a region tracks a state does not tell you whether stimulating that region will produce the state, suppress it, or do something unrelated. That is exactly what the interference experiments on working memory showed: stimulating the neurons that carry the representation degraded the representation rather than reinforced it.

The system also adapts. The brain is plastic: it changes in response to the stimulation you apply, so a controller calibrated today can be miscalibrated tomorrow as the system it learned to steer becomes a slightly different system. Detection can track a drifting signal in real time; writing must act on a moving target whose future behavior is only partly predictable.

Put together, these facts make writing a control problem in the strict engineering sense, and control problems in biology are hard for well-understood reasons. The plant is nonlinear. Its response depends on its state. The sensor measures a proxy, not the target quantity. Disturbances are constant. And the objective — what state, in which person, toward what end — is not given by the physics at all.

What writing to a brain demands

If writing is this hard, what would count as having done it well? The honest bar is stricter than a striking demonstration. A credible writing capability would need to show not only that a target was engaged but that a specific, intended functional change followed — reliably, in more than one person, with the opposite effect demonstrably not occurring. It would need to report side effects and off-target effects rather than only the target effect, because a tool that reaches its goal by perturbing everything nearby has not learned to write to a circuit. It would need to be durable enough to matter clinically and reversible enough to be safe. And it would need to establish that stimulation parameters transfer across individuals, or that personalization can be achieved cheaply, since the fragility in the studies above comes largely from the fact that optimal settings differ from person to person.

No current method clears that bar for anything as complex as a mood or a cognitive mode. The narrow successes — a pulse that disrupts an ongoing seizure, a stimulation protocol that improves memory in a controlled task, an electrode that relieves a movement symptom — are real and, in several cases, deeply valuable. They are also bounded, and they cluster around disruption and symptom relief rather than around the precise, selective writing that the popular phrase “programming the brain” implies.

Where the asymmetry leads

The practical upshot is that the technology to read a neural state will almost certainly mature ahead of the technology to write one, and the two should never be conflated in thought or in policy. That gap has consequences for the whole conversation about neural technology. A decoder that can estimate a person’s state, paired with only crude tools for changing it, is not a mind-control device; it is a monitoring instrument with a blunt actuator attached. The danger in such a system is less about covert control than about what can be inferred, stored, and acted upon outside the person’s awareness — the questions of data and agency that run through the rest of this series.

As the writing tools improve, the asymmetry narrows, and the stakes rise with it. A tool that can reliably and selectively change a neural state is the most consequential thing this field could build, precisely because the whole difficulty described here — the fragility, the state-dependence, the reach through a degrading skull — is also what currently keeps the power small. The work of getting governance right is the work of preparing for a capability that is approaching unevenly, arriving first as reading, and only later, if at all, as writing.