The Programmable Emotional Landscape

A person recovering from a severe depression often describes the change in dimensional terms rather than with the name of an emotion. The flatness lifted, they say; colors returned; the volume on threat came down. They rarely report that a single mood was installed. If a technology eventually tunes how a person feels, it will probably work the same way — not by selecting “happy” from a menu but by moving one or more underlying dimensions while leaving the rest of the system to respond. Whether that is possible, and what it would mean, is a question about the structure of emotion itself, not just about the reach of our tools.

The popular picture of emotion as a set of discrete buttons — fear, anger, joy, sadness — has a long history but an increasingly contested empirical footing. The alternative, a dimensional model in which feelings vary along continuous axes, turns out to describe both subjective experience and some aspects of brain activity better, and it also makes the engineering question sharper. Tuning a dimension is a different act from suppressing an emotion, with different ethics and different failure modes.

Emotion as coordinates rather than categories

The dimensional view was formalized by James Russell in a 1980 paper in the Journal of Personality and Social Psychology that proposed what he called a circumplex model of affect (Russell, 1980). Russell found that emotion words, when people rated them for similarity, arranged themselves around a circle rather than clustering into a few independent families. The organizing axes were pleasure versus displeasure and arousal versus sleepiness. “Excitement” sat in one quadrant, “contentment” in another, “distress” in a third, “relaxation” in a fourth. The labels were points on a surface, and the surface was two-dimensional.

Russell and colleagues extended the framework to affective neuroscience in a 2005 review in Development and Psychopathology, arguing that the two axes — valence and arousal — arise from largely independent neurophysiological systems and that a specific emotion such as fear emerges only when those core dimensions are combined with cognitive interpretation of the situation (Posner, Russell & Peterson, 2005). This is a substantive claim, not a relabeling. On the basic-emotion account, fear is a dedicated program with its own circuitry. On the dimensional account, fear is a construction placed on a moving point in valence-and-arousal space, and the same point could be read as excitement or as dread depending on context.

Imaging work has given the dimensional model partial support while also exposing the fragility of the basic-emotion story. In a 2009 study in Human Brain Mapping, Jonathan Posner and colleagues had participants process emotion-denoting words during fMRI and found that valence and arousal correlated with distinct patterns of brain activity, consistent with separable systems (Posner et al., 2009). The same paper contains a fact worth pausing over: across two meta-analyses, amygdala activation during fear was reported in only about sixty percent and forty percent of studies respectively. If fear had a dedicated, reliably firing neural signature, that number should be closer to all of them. The inconsistency is a reminder that a great deal of confident public talk about “the fear center” rests on a shakier imaging literature than the talk implies.

What has actually been decoded and what has actually been changed

The dimensional model does not by itself imply that emotion can be read or written. Two very different lines of research test those claims, and they sit at opposite ends of the demonstrated-to-speculative spectrum.

The reading side has a genuine, if narrow, result. In a 2018 paper in Nature Biotechnology, Omid Sani, Yuxiao Yang, Maryam Shanechi and colleagues recorded from intracranial electrodes implanted in seven patients being evaluated for epilepsy surgery and had them repeatedly report their mood using a validated questionnaire (Sani et al., 2018). Mood could be decoded from the multi-site brain activity, but the decoders were individual: a model that worked for one patient did not transfer to another. The electrodes that carried the most information sat in limbic regions, and the authors were careful to describe the work as an initial line of evidence rather than a general capability. Intracranial electrodes are placed for clinical reasons, in a small number of patients, and decoding mood from them says nothing about reading emotion from a consumer wearable.

The writing side has dramatic animal results that are frequently misunderstood. In a 2023 study in Neuron, Piantadosi and colleagues used two-photon holographic photostimulation in mice to activate specific ensembles of neurons in the basolateral amygdala that had been associated with either a positive or a negative response to taste (Piantadosi et al., 2023). Stimulating the “appetitive” ensemble made the animals consume more sweet solution; stimulating the “aversive” ensemble made them reject it. The two ensembles inhibited each other, and the behavioral outcome depended on the proportion of neurons recruited. This is a causal demonstration that valence-related behavior can be switched by manipulating a defined neural population — in mice, through an implanted lens, in a controlled laboratory setting.

A companion line of work clarifies how much of that manipulation depends on where you intervene. In a 2014 paper in Nature, Redondo, Kim, Arons, Ramirez, Liu and Tonegawa showed that a memory engram in the dorsal dentate gyrus could have its valence switched: activating the same cells that had encoded a negative experience could produce approach behavior instead of avoidance (Redondo et al., 2014). In the same set of experiments, the valence attached to an engram in the basolateral amygdala could not be switched. The finding cuts against the idea that emotion is uniformly plastic. In some circuits the sign is settable; in others it appears hard-wired. A technology that promised to retune emotional valence everywhere would be making a claim that this evidence does not support.

The clinical evidence is more sober than the headlines

If emotional dimensions can be tuned, the most likely place to see it first is in treatment, where the manipulation is already happening with drugs. Two recent cases show how differently that promise has played out.

Compass Pathways reported top-line results in June 2025 from the first phase 3 trial of its synthetic psilocybin, COMP360, in treatment-resistant depression, and the company’s release and its filing with the SEC give the numbers plainly (Compass Pathways, 2025). A single 25 mg dose beat placebo on the MADRS depression scale by a mean difference of 3.6 points, with a confidence interval from 5.7 to 1.5. The trial enrolled 258 participants across 32 US sites, and its safety monitoring board reported no new safety concerns. The result is statistically solid and clinically modest. A 3.6-point advantage is real, but it is a fraction of what a person with severe depression would call recovery, and the endpoint was measured at a single early time point. The FDA had not yet reviewed the data when the company announced it.

The MDMA case is the cautionary one. In August 2024 the FDA issued a complete response letter declining to approve midomafetamine capsules for post-traumatic stress disorder, and the reasons are instructive for anyone imagining emotional technology as a clean therapeutic path (FDA, 2024). The agency found that the application did not provide substantial evidence of effectiveness and did not establish safety. Among the specific concerns were unreported adverse events described as positive or euphoric, adverse events that had not been reported at some trial sites, and a study population in which roughly forty percent had prior MDMA use — a pattern that raises the possibility of functional unblinding, since participants may have guessed whether they received the drug. The agency also questioned whether the treatment effect was durable and asked the sponsor to conduct another phase 3 trial. The lesson is not that psychedelic-assisted therapy cannot work. It is that a treatment which produces strong subjective experiences is unusually hard to test, because the experience itself can break the blind on which causal inference depends.

Where the boundary between demonstrated and speculative falls

Drawn carefully, the map looks like this.

Demonstrated: emotions can be represented along continuous dimensions with valence and arousal as the principal axes; mood-related brain activity can be decoded from intracranial recordings in individual patients; valence-related behavior in rodents can be switched by activating defined neural ensembles; and several treatments, from psilocybin to implanted electrodes, produce measurable changes in depression severity under controlled conditions. All of this is real and replicated, within the stated limits.

Plausible engineering: closed-loop systems that read a person’s state from implanted or, in the future, high-quality surface signals and adjust stimulation accordingly; psychotherapy or pharmacology guided by continuous, individualized mood tracking; and the tuning of one dimension — arousal, say, or threat reactivity — without attempting to control the whole emotional landscape at once.

Speculative: a general, non-invasive emotion dial for healthy people; decoding that generalizes across individuals from a wrist-worn device; and any claim that a specific emotion can be reliably written into a person the way a parameter is set. The evidence does not currently support the first two, and the finding that some amygdala valence signals are not switchable directly challenges the third.

The difference between tuning a dimension and muting a signal

Suppose the engineering matures. The ethical question does not resolve, because tuning a dimension and suppressing a symptom are different operations even when the numerical change looks identical.

Consider someone grieving a death. Grief is unpleasant, and by the dimensional model it sits in a particular region of valence and arousal. A system told to reduce negative valence would move them out of that region. The trouble is that grief is also doing work — it is how a person’s relationship to what was lost gets reorganized over time, and it carries information that the loss was real and mattered. A tool that flattened the grief would not be treating a disorder; it would be deleting a response that the situation calls for. The same logic applies to anger at an injustice, guilt over a genuine wrong, and the anxiety that accompanies real danger. An emotional state can be both aversive and correct, and nothing in the technology can make that distinction.

One response is a design principle rather than a prohibition. Systems aimed at recovery of function — helping someone eat, sleep, work, and re-engage with people — aim at a target that the person can endorse independent of how they feel in the moment. Systems aimed at smoothing emotion aim at a target defined by the absence of distress, which is easier to specify and easier to pursue past the point where it helps. The clinical literature already reflects this distinction: depression severity scales measure impairment in functioning, not the mere presence of sadness, because sadness alone is not the disease.

The problem of the moving target

There is a second difficulty, particular to emotion, that does not arise as sharply in the sleep case. When a person’s preferences are part of what is being changed, the baseline against which consent was given can drift out from under the decision.

A person may agree to a course of treatment while depressed and later, having improved, disagree with the decision the depressed version of themselves endorsed. This is not a hypothetical; it is a routine feature of psychiatric care, and it is why clinicians work with advance directives and with the distinction between a person’s stable values and their symptoms. An adaptive system that continuously retunes emotional dimensions makes the problem worse, because the “target” itself may be recomputed against a state the person has since left. Any such system needs a way to hold certain commitments fixed — to ask the person who is not currently being tuned what they want the system to be allowed to do.

Everybody already tunes emotion, just crudely

It would be a mistake to treat emotional manipulation as a new phenomenon that technology introduces. Music, caffeine, exercise, sleep, alcohol, social contact, religious practice, and the deliberate reframing that therapists teach all change how a person feels along valence and arousal. A runner’s post-exercise calm, a strong coffee’s lift, and the numbing of a second drink are all, in the dimensional vocabulary, movements within the same space the brain already occupies.

What changes with more capable technology is precision, speed, and the possibility of feedback. A song you chose is different from a system that observes your state, predicts where it is heading, and adjusts an input you are not consciously tracking. The distinction that the rest of this series keeps returning to applies here as well: does the tool widen the range of states a person can voluntarily enter and leave, or does it narrow that range while making the narrowing feel natural?

Emotion is a poor candidate for the fantasy of a permanent setting, because its variation is part of its function. A person locked into one region of the circumplex is not a person who has solved emotion; they are a person who has lost the ability to respond flexibly to a changing world. The most defensible versions of this technology will aim at restoring that flexibility where illness or injury has removed it, and will treat the ability to feel bad when circumstances warrant it as something to preserve.

What would settle the question

Several specific results would move this from speculation toward a settled capability, and their absence is as informative as their eventual arrival would be.

A decoder that works across people without individual calibration, in a non-invasive device, would establish that emotion can be read reliably rather than inferred from behavior. A stimulation method that changes a dimensional measure and the corresponding behavior and holds under double-blind conditions would establish that it can be written. A clinical trial that shows a durable benefit after the intervention is withdrawn, with adverse events reported including the ones that flatter the treatment, would establish that the change is therapeutic rather than transient. And evidence that a manipulation preserves appropriate negative emotion while removing pathological intensity would establish that the technology can distinguish a symptom from a signal — which is the distinction the whole question turns on.

Until that evidence exists, the honest description of the programmable emotional landscape is a map with a well-surveyed coastline and a mostly blank interior. The coastline — dimensional structure, individual decoding, causal manipulation in animals — is real. The interior is where the promises live.