What an Emotion Actually Is

An emotion is an error signal in a feedback loop. That single claim, borrowed from cybernetics and control theory, changes what hunger, fear and shame are for. It also explains why happiness refuses to behave like any of them.

Ask a room of clinicians what an emotion is and you'll get answers about valence, arousal, appraisal, and evolved function. All useful. None of them tell you what an emotion is for in the way an engineer means the word.

I came at this sideways. Before I was a licensed psychotherapist I was an aerospace engineer working in guidance, navigation and control. Feedback loops, with an estimate on one end and a correction on the other, and a short list of ways they fail. That vocabulary transferred into the therapy room better than it had any right to.

This is the first of seven pieces working through a model that joins control theory to predictive processing. It starts here, with the half that psychology mostly abandoned.

What is an emotion, mechanically?

Error doesn't mean bad here. It's a signal that something is off, and it arrives as a feeling rather than as a number.

The vocabulary is small enough to carry around.

A word on cybernetics, since it's been dragged off in the wrong direction. It has nothing to do with cyborgs and it predates them. It comes from the Greek kybernetes, the steersman on a ship, and Norbert Wiener took it in 1947 for the study of how any system steers itself by feedback (Wiener, 1948). A man at a tiller. A thermostat. A cruise control. A body holding its temperature.

Wiener chose that word because the first serious paper on the subject was Maxwell's 1868 study of governors, the spinning weights that hold a steam engine at speed (Maxwell, 1868).

The clearest recent statement of the emotion claim is a self-published series called The Mind in the Wheel, by a group writing as Slime Mold Time Mold. Their argument, across a prologue and twelve parts, is that psychology never had a paradigm because it never proposed actual entities and rules, only abstractions.

Their candidate entity is the negative feedback loop, which they call a governor: a sensor, a set point, a comparator, and an output that acts on the world. Hunger, thirst, fear and shame are error signals from different governors, all competing for one body through an arbitration layer they call the selector.

From there they rebuild personality as the parameter settings on your governors, and recast depression not as one disease but as several mechanically distinct faults that happen to share a surface. It's self-published, it has essentially no new data, and it's the most complete recent statement of a control-theoretic psychology anyone has written.

None of this is new to psychology, which is part of why it's worth taking seriously. William Powers built an entire theory of behavior on the claim that organisms control their perceptions rather than their actions (Powers, 1973). Carver and Scheier carried feedback control into personality and clinical psychology forty years ago (Carver & Scheier, 1982). The idea has been sitting there, on the losing side of an argument, since the cognitive revolution.

Why isn't happiness an emotion?

It's the first question anybody asks, and it's a good one.

The Slime Mold answer is direct:

"Emotions are easy to identify because they are errors in a control system. Like any error in a control system, successful behavior drives the error to zero. This means that happiness is not an emotion."

What happiness is, on their account, is what it feels like when a governor's error gets corrected. Their line: "Happiness is what happens when a thirsty person drinks, when a tired person rests, when a frightened person reaches safety."

Correct a large error, or correct one quickly, and you get more of it than from a slow incremental fix. That matches ordinary experience better than it has any right to. The best meal of your life was eaten hungry.

It also predicts something odd that I think is right. There's no such thing as unhappiness. Happiness can't go negative. What gets called unhappiness is an uncorrected error somewhere, which is a different thing wearing the same word.

And they give it a job. Happiness marks which behaviors worked so the system repeats them, and it helps calibrate how much to explore versus stay with what already works.

If the list of regulated variables reminds you of Maslow, the resemblance is real but the logic runs the other way. Maslow arranged needs into a hierarchy of priority. This arranges them as parallel loops competing for one body, which is why a person can be hungry and lonely and cold at the same time without any of it queuing politely.

What does this model leave out?

I'd rather leave those unaccounted for than stretch the framework to cover them.

The Slime Mold position is wider than mine. They handle joy, relief, satisfaction and pleasure by making them the same thing. Their words: "there's only one way to feel good," and "all of our words for positive emotion, joy, excitement, pride, are really referring to the same thing, just in different contexts."

That's asserted rather than argued, and it leaves amusement out entirely. Across fourteen posts, amusement, awe and aesthetic pleasure don't turn up anywhere.

They're straighter about curiosity, which they flag themselves as an enigma: "acting on your fear should make you less afraid, acting on your thirst should make you less thirsty, but acting on your curiosity often seems to make you more curious." A signal that grows when you act on it is not behaving like an error.

Stretching a model until it covers everything is how a model stops saying anything. Part 7 of this series is about exactly that failure, and it names the results that would show this whole framework is wrong.

How does a control loop break?

The pair that matters most is gain and damping.

A loop with high gain and insufficient damping doesn't just respond too strongly. It oscillates. It overshoots, corrects, overshoots the other way, and either settles slowly or never settles at all.

Wiener identified cerebellar tremor as exactly this problem in 1948. Somebody with cerebellar damage reaches for a glass and the hand doesn't travel smoothly. It overshoots, comes back too far, overshoots again, and hunts around the target, getting worse the closer it gets.

His point was that this isn't weakness and it isn't paralysis. The muscles work fine. What's broken is the correction, which arrives too hard and too late. That's the textbook signature of a feedback loop with its damping removed.

Hold onto that pattern. It reappears in rumination, in compulsions, and in a manager who checks on a project every two hours.

What happens when a person replaces the thermostat?

Work the example all the way through, because it's the cleanest way to see why a controller alone isn't enough.

Your wall thermostat already has predictions built into it. Not obvious ones. It waits before it adjusts. That's hysteresis, meaning the point where it switches on isn't the point where it switches off. Once the air conditioning is on it holds until the temperature passes the set point rather than flipping the instant the number crosses. None of that is comfort engineering. It's there so the unit isn't cycling on and off on sensor noise.

Now take it off the wall and put a person in its place.

Here's what they don't know at the beginning. There's a delay between turning that dial and their body registering the change. The unit has to start, the air has to circulate, the room has to cool, and only then does their body notice.

They crank it up when they're hot, nothing happens, so they crank it further. Then the room goes cold and they crank it way down. Then it's hot again. Each correction is bigger than the last, because they aren't getting the response they expect on the timeline they expect it.

Now let them learn. Once they know about the delay the behavior changes completely: crank it up a little, wait, then decide whether to adjust again. Same dial, same body, same room. What changed is the prediction, and the prediction is what damped the controller.

That splits the job in two, and the split is the whole point of this series.

Two different failures live in those two jobs, and from outside they look almost identical.

Go back to the dial. Suppose the sensor is bad. It reads 70, then 71, then 68, then 71, and the system reacts to every reading. The air conditioning cycles constantly, and nothing at all is wrong with the decision rule. The reading is the problem.

Or suppose the sensor is perfect and the patience is missing. It's exactly 70.0. At 70.1 it turns on, at 69.9 it turns off. Perfect information, no prediction, same cycling.

It isn't all failure, either. Put your hand on a knife and cut yourself. The pain arrives large and the response is immediate, which is the controller doing its job correctly, because fast and hard is the right answer to a knife.

Meanwhile the other half of the system is doing something slower and more useful. It takes that sensory input against what you expected and learns from it: that hurt because the knife is sharp. You now carry a prior that knives are sharp, and your caution around them is set at a more accurate level than it was an hour ago.

One event, two jobs. The controller handled the moment. The estimator changed the setting.

Which is also why the same parameter is right in one loop and wrong in another. For pain and danger, act fast. For whether the room is a little warm, you can afford to be slow. React to discomfort on the timescale you'd react to danger and you're the person cranking the dial.

Where does the control model run out?

That's a hole where the sensor ought to be, and it fits almost exactly what predictive processing has spent thirty years building.

Which is where Part 2 goes.

References