Brains Are Body Regulators: Lisa Feldman Barrett on Why Emotions Are Constructed, Not Detected

Lex Fridman introduces Lisa Feldman Barrett, a Northeastern University psychologist and author of *How Emotions Are Made*, and frames the talk as a way to gain intuition about instilling emotional life in AI. After a warm-up about films that move people to tears, she names the most common misconception: that you can read someone's emotion from their face as you read words on a page.

Barrett says many tech companies pour money into building emotion-detection systems when they are really building excellent systems for reading facial movements. People scowl when concentrating, sometimes when sad or even happy, and they smile when angry or sad. She says there's no strong scientific evidence for one universal facial expression per emotion.

She adds that emotions aren't simply displayed: when you look at someone, your brain takes in their voice, posture and the dynamic changes across all of these signals, even though it feels like all the information is in the face. She jokes about immigrants whose cheeks ached from smiling in the US. The face, she says, is not a lighthouse that broadcasts an internal state.

Barrett says scientists have argued for millennia over two views. In one, everyone is born with dedicated circuits for anger, fear, sadness and happiness that trigger obligatory, stereotyped responses in the body, face and action. In the other, which she supports, the brain has common ingredients and assembles emotions on the spot as needed.

She uses a kitchen analogy: flour, water and salt can make many recipes, even glue. On this view you have a whole repertoire of anger or sadness rather than one of each, and a culture with no concept of sadness would not produce the experience. She cautions that she wouldn't call it a lack of an answer from science, only that scientists disagree.

For AI, she says, emotion is the wrong thing to try to endow. Some cultures have no concept of emotion, but everyone has simple feelings of pleasant, unpleasant, worked up or calm, which she calls affect, driven by the body's internal state. What an agent needs are the basic ingredients to build whatever states and actions suit its environment, which differs between a Western city and a hunter-gatherer group in Tanzania.

Barrett argues that brains didn't evolve so people could think, feel and see; they evolved to control bodies as creatures move around gaining resources such as food and water. A brain must decide how many resources to spend to get more, which sounds trivial but is hard, and scientists haven't fully figured out how brains do it. She also rejects the popular lizard-brain, or triune brain, story of evolution.

To illustrate, she says computer vision is largely solved at a basic level, while making an agent reach out smoothly, grab a glass and bring it to its body is not. Movement, which seems so basic, is among the hardest problems.

Parts of your brain are always controlling your heart, lungs and immune system, and they also summarize the sensory consequences of those changes. You usually don't feel your heartbeat, because you'd otherwise never pay attention to the world, so the brain produces simple summary feelings instead. These feelings are with you every waking moment and are not emotions, but sometimes the brain turns big changes into emotion, hunger, nausea or even a perception such as a delicious drink.

Fridman asks whether saying emotions aren't built into the brain means they aren't real. Barrett corrects the inference: she isn't claiming people don't express or feel emotion. She points out that heart rate or breathing can go up, down or stay the same in anger, and similar patterns show up in sadness, fear, awe and enthusiasm.

Her distinction is between information that is in a signal and meaning the brain imposes on it. A retina's wavelength is in the signal, but the color red is added by the brain. Many social things are real by collective agreement, such as money, countries and presidents; a revolution is people withdrawing consent, and the 17th-century tulip crisis and mortgage crisis show value vanishing when agreement does.

Emotion works the same way. We impose meaning on physical signals, and because we all agree that scowling is sometimes anger, it becomes anger in a very real way. Fridman likens it to a language, and she agrees; people who move cultures have to learn a new social reality, and failing to do so can make them physically ill.

Asked how people learn to express emotion, Barrett starts with infants, who cannot regulate their own nervous systems and rely on caregivers. She defines allostasis as the brain predicting what the body will need before it needs it, such as raising blood pressure before standing. Infants are doing statistical learning from the start, linking sights, sounds and body sensations to whether needs are met, and they learn faces in the first days because faces share regularities.

Around three months, she says, infants begin forming abstract categories, learning that very different-looking objects share a function when they share a word. She describes experiments where an infant told that objects are the same word expects them to behave alike. Words are invitations to form categories, and that is the basis of money, rules and mental categories like emotion concepts.

She portrays the brain as living in a dark, silent box, getting only the effects of events in the world and body. A flash, a siren or a stomach ache could have many causes, so it uses past experience to guess, and the information it did not predict is what psychology calls learning. Even an ache in the stomach could be hunger, nausea or longing, she says, so the brain makes the same sort of guesses about the body as about the world.

Asked how to build a robot that could pass for human, Barrett says it needs something like a body and something analogous to allostasis. She argues that in machine learning reward is treated as simple math, but in biology it's tied to bringing the body back into balance, with dopamine more about the effort needed to pursue a reward than reward itself. Motivation, she says, is spending resources to get reward, and the need to keep water, salt and glucose in balance is the motivating force.

On whether a robot could fall in love, she jokes that people fall in love with cars and blankets, and that the harder question is whether it loves you back. She describes humans as social animals who regulate each other's nervous systems using touch, smell, vision and, uniquely, words, which connect directly to brainstem systems controlling the body. That is why a voice on the phone can slow your breathing.

Because brains are metabolically costly, taking about 20 percent of the body's budget and limited in size, she says evolution enlisted other brains to help manage each person's nervous system. She says attachment is the ability to affect each other's nervous systems positively, and loneliness kills on average seven years sooner. Her line: the best thing for a human nervous system is another human, and the worst is also another human.

In audience Q&A, a student asks how novelty fits with returning to balance. Barrett answers that brains aim for metabolic efficiency, not perfect prediction, since perfect prediction would be boring, so there is a balance between exploiting and exploring. Novelty feels exciting but is expensive because it involves learning something new and moving the body.

She compares two breeds of rats in a new cage: one sits still, one roams everywhere. The still rat is not necessarily afraid; it may be saving, while the other is spending and encoding a lot. She likens the brain to a company's finance office, noting that when a deficit appears you can either take a risk to earn big or cut spending.

How readily any nervous system embraces novelty depends on its allostatic state, she says, and continuing to spend while in the red leads to bankruptcy. In human terms that is depression, where the brain fatigues you, stops you moving and locks you into what's in your head. She also describes allostasis as a range rather than a single set point.

Asked whether general intelligence needs to be embodied, Barrett says to be careful: biology has degeneracy, meaning many ways to reach the same function. So an agent needs something akin to a body, perhaps multiple systems it must manage, rather than a literal one. She points out that a brain needs only to have once had a body, citing phantom limb and chronic pain, where the brain simulates sensory patterns from a body that no longer reports.

A student suggests emotions are evolutionary hacks involving chemicals like serotonin and oxytocin that AI could do better without. Barrett rejects the premise: no chemical is for emotion. Serotonin, for example, helps the brain delay gratification and spend energy now for a future reward, and opioids and dopamine do their work in every waking moment, not only emotional ones.

She repeats that from a cybernetics view the best way to regulate a system is to build an internal model of it, and that the brain is an internal model of the body in the world, running simulations. Whether that needs a physical body is, she says, an open question.

A student asks how to make a system that desperately wants to create something new. Barrett says a body is a necessary condition, not a sufficient one, and that what people find rewarding is diverse: many in the room are driven by impact or building something new, but not everybody is. She adds nothing she said conflicts with the evolutionary idea of passing on genes, and that learned social concepts are a cheap way to wire the young brain with what a society knows.

On intentionality, she warns that the word means both deliberate action and having a referent in the world, and that the brain is always making choices even when we don't consciously experience agency. Humans are not sea urchins, she says; interneurons mean we have choice, but we don't experience much of it as deliberate.

Finally, asked about dogs, she says many animals have affect, and dogs are special because humans bred them. They have fine control of their facial muscles and use shared gaze to refer to things, an ability chimpanzees lose after about a year. Shared gaze is also how humans teach infants what matters in the world.