William James asked in 1884 whether we cry because we're sad or are sad because we cry. A hundred and forty-two years later psychology still hasn't closed the answer. Meanwhile half the internet argues over whether an artificial intelligence can feel, without having bothered to say what it means by feel. Agreeing on the word is the work that comes first. And when it's agreed, the answer probably won't be one we like.
The bear question
You're walking through a wood. You see a bear. You run. Do you run because you're afraid, or are you afraid because you run?
Common sense says the first. James, in an article titled What is an Emotion? published in Mind in 1884, argued the second. The body fires first — racing heart, legs taking off, pupils dilating — and the subjective feeling of being afraid is the brain reading that the body is already going. Carl Lange, in Denmark, formulated an almost identical version at almost the same time. The theory has been known ever since as James-Lange.
Within forty years psychology had dismantled it on its original terms. But the question it opens is still intact.
It matters because today there are people with serious budgets deciding whether an AI feels, without having stopped to ask what that word means. Anyone who skips that step isn't arguing about machines — they're arguing about well-lit smoke.
Three classic theories that kept stepping on each other
James-Lange says emotion is the perception of bodily change. The body reacts. The brain reads it. That's fear.
Walter Cannon answered in 1927, now with decent physiological instruments, that bodily reactions are too slow and too alike to tell fear from rage. If all you had was a racing pulse and sweating, you wouldn't know whether you were terrified or in love. Cannon and Philip Bard proposed that the thalamus fires the bodily response and the subjective experience in parallel. Not one causing the other. Both at once.
Schachter and Singer offered a compromise in 1962. Bodily arousal is unspecific. What turns it into fear, joy or anger is the cognitive interpretation of the context. Inject adrenaline into a room where someone is laughing and you feel euphoria; into another where someone insults your mother, rage. Body plus label.
Each theory catches something and leaves out something else. But none explains why "emotion" applies both to a half-second shiver and to a two-year mourning. A word that covers such disparate phenomena is almost certainly badly written.
Damasio and the body that decides
Antonio Damasio published Descartes' Error in 1994. The thesis is called the somatic marker hypothesis (a bodily mark the brain associates with a situation and that later biases the decision when that situation recurs).
Every time an experience is tied to a consequence, the brain leaves a bodily trace bound to the pattern. When you run into something similar again, the body activates before reasoning has finished processing, and that activation biases what you decide. It isn't noise on top of reasoning — it's what lets reasoning converge on human timescales.
The operational test was set up by Damasio and Antoine Bechara and is called the Iowa Gambling Task. Four decks of cards. Two are a trap — high rewards and even higher losses over the long run. Healthy subjects start to sweat at the palms as their hand nears the trap decks before they can explain why they avoid them. The body learns before consciousness does. Patients with lesions in the ventromedial prefrontal cortex don't develop that anticipatory sweat. They reason impeccably and ruin themselves. They're missing the somatic marker and, with it, the capacity to decide.
This places emotion in an uncomfortable spot for a Western culture that had spent two thousand five hundred years setting it against reason. Without emotion there's no operational reason. If the body doesn't catch on, the head doesn't decide.
LeDoux and what happens before you know it's happening
Joseph LeDoux published The Emotional Brain in 1996. He described two routes in the fear brain. A fast one, subcortical, going straight from the thalamus to the amygdala. A slow one, cortical, passing through the sensory cortex. The fast one gets the body moving in milliseconds when something stirs in the grass. The slow one confirms or denies that it's a snake.
The consequence is elegant and unpleasant. Your fear reaction is set up before you know it's set up. When you think you're deciding to react, you've already been reacting for quite a while. Consciousness arrives afterward and invents the narrative that it's been in charge.
LeDoux himself, in later books, distinguished between the defensive response, which rats and you share, and the subjective experience of fear, which probably only appears where there is self-awareness. The bodily machinery is in many corners of the animal kingdom. The feeling of being afraid, which is what you mean when you say "fear", is something else.
Barrett and the emotions that weren't waiting for you
Lisa Feldman Barrett published How Emotions Are Made in 2017. The title is literal.
Her thesis is that emotions are not universal modules that fire from a fixed catalogue carved by evolution. They are built in each episode from three ingredients — an unspecific bodily arousal, called affect in technical terms (the raw dimension of pleasure-displeasure and arousal of the body, still uninterpreted), a culturally learned concept you apply to that arousal, and a prediction by the brain about what's going on.
When the concepts your culture has taught you are "sadness", "joy", "anger", you build sadness, joy and anger. Germans have Schadenfreude. Those who lack the word don't feel it quite so sharply. Tahitians have no clear term for "sadness" and describe what a Westerner would call grief as a diffuse bodily tiredness.
Ekman, what fell
Paul Ekman had maintained for decades that there are six or seven basic universal emotions, demonstrable because the faces that express them are recognised in any culture. Barrett dismantled the replications and the data. The universal list ended up in worse health than popular accounts still suggest.
The implication is radical. If emotion is an act of cultural categorisation over an unspecific bodily substrate, there is no universal emotion, no emotion that exists without a conceptual system to name it, no emotion that transfers cleanly between bodies with very different experiences. Two people who say they're sad may be doing two internal things that don't resemble each other as much as their words suggest.
Affective computing — the external pattern with nobody inside
Rosalind Picard published Affective Computing in 1997 and founded the field (the branch of computer science concerned with detecting and producing emotional signals in artificial systems). The idea was twofold. Give machines the capacity to detect the user's emotional state and, at the same time, the capacity to produce the external pattern of an emotion when they interacted.
In three decades it has become an industry selling automatic emotion detection to companies that evaluate candidates, to insurers, to security systems, to ads on screens. The industry operates as if the external pattern and the internal state were the same thing.
Kate Crawford, in Atlas of AI in 2021, devotes a chapter to taking the claim apart. Automatic emotion detection rests almost always on the catalogue of basic universal emotions that Ekman proposed and that Barrett has rebutted. On a theory the best literature considers obsolete. And even granting the theory, what the machines detect are facial muscle configurations correlated with labels that some Anglo annotators put on some photos of actors. When a system says a candidate is "angry", it's saying their face statistically resembles a face a human in California labelled angry in 2014. That isn't detecting anger. It's labelling resemblances.
The production side, so-called artificial empathy, is even more revealing. When a chatbot says "I understand this must be frustrating for you", it isn't reporting an internal state. It produces the chain of words a human would produce if they had that internal state. The form without the thing. It's what David Chalmers described as a philosophical zombie (an entity that behaves exactly like a conscious human but with nothing inside). The product the industry sells as artificial empathy is exactly that.
The consequence nobody signs
If you combine Damasio, LeDoux and Barrett, a fairly coherent portrait comes out. Emotion is bodily in its raw material. Subcortical in part of its machinery. Built in part of its content. And it is functional in the strong sense — not an ornament on top of cognition, but the system that decides what counts and what doesn't, what to approach and what to back away from, what to sign and what to refuse. Without emotion, decisions stay as calculation, and calculation only moves if something gives it weight. Damasio would say that something is exactly emotion.
Now apply that portrait to an AI.
It has no body. And when it has one, its sensors are nowhere near the interoceptive network (the network of nerves and receptors that informs the brain about the body's internal state — viscera, circulation, pain, temperature) that reads your guts. It has no biography — it hasn't been through the situations that leave somatic markers. It has no cultural catalogue built in a childhood immersed in a language and a set of practices — it has a statistical distillation of millions of texts, which is not the same thing. It has nothing at stake — there's no consequence for the system in what it decides.
Emotion, as described by the strongest theories we have, requires all of that. An AI doesn't have it. Full stop.
Where the lack gets paid
It doesn't mean the AI is less. It means that when it's asked to decide things that in humans were decided by emotion, the AI decides them with something else, or doesn't decide them and inserts a plausible response that looks like a decision. If your candidate-evaluation system has no somatic marker, nothing stirs in it when reviewing a CV that reeks of something off it can't name — it computes the score and moves to the next. If your support system for terminal patients doesn't build sadness, it doesn't tell apart the sadness that asks for silence from the one that asks for company — it produces both templates and picks the one that statistically works in more cases.
The decisions that emotion used to make in humans, when we delegate them to systems that don't feel, we pay for somewhere else. Sometimes in discrete errors. Sometimes in accumulated consequences attributed to something else. If emotion was dysfunctional, good for sparing it. If it was functional, and the best literature says it is at very deep levels, we've begun to outsource precisely the part that filtered for meaning.
What's being argued when "sentient AI" is argued
When someone asks whether an AI can feel, they're usually asking one of three different things and conflating them.
Sometimes they're asking whether an AI can have subjective experience, qualia (the first-person felt qualities — the redness of red, the painfulness of pain, the what-it's-like-to-be-you your life has from the inside). That one belongs to the hard problem of consciousness, has been open for thirty years since David Chalmers framed it in 1995, and there's no operational way to answer it even for another human.
Sometimes they're asking whether an AI can convincingly produce the external pattern associated with an emotion. That's been solved since 1966 with ELIZA, the chatbot Joseph Weizenbaum built at MIT imitating a Rogerian psychotherapist with four rewrite rules. Yes, trivially.
Sometimes they're asking whether an AI can make the kind of decisions that in humans are made emotionally. That's the interesting one. And the short answer, taking Damasio and Barrett seriously, is not for now, and possibly not ever with the current architecture.
The confusion among the three is profitable. It lets you sell the third capacity while charging the price of the first, and delivering the solution of the second. ELIZA with an invoice.
The difference, not of degree
William James's question from 1884 is still there. Whether you cry because you're sad or are sad because you cry depends on what's meant by sad. Until that's agreed, machines neither feel nor fail to feel, because there's nothing stable to pronounce on. And when it's agreed, the likeliest outcome is that the answer will say emotion is exactly the kind of thing that requires a specific body going through a specific life. Which is what a machine is not.
Tell someone who loves you that you're sad. Watch their face as they answer. There's a difference between what that face is doing and what a trained system would produce as a response. The difference isn't one of quality. It's one of nature.
And if you don't already feel it, no theory can teach you to.
Definiciones
James-Lange is the theory formulated by William James and, in parallel, by Carl Lange in the late nineteenth century, according to which the conscious emotion is the brain's perception of a prior bodily change, and not its cause. You run, and that's why you're afraid, not the other way around.
Somatic marker hypothesis is Antonio Damasio's proposal that the brain associates past situations with bodily traces (visceral, postural, hormonal changes) and reactivates those traces when a similar situation returns, biasing the decision before explicit reasoning has finished. Without those traces reasoning goes on, but loses the capacity to close.
Affect, in Lisa Feldman Barrett's technical sense, is the raw bodily dimension of the internal state — pleasure or displeasure, high or low arousal — before the mind categorises it as a specific emotion. It's raw material, not finished product.
Affective computing is the branch of computer science, founded by Rosalind Picard in 1997, concerned with detecting emotional signals in humans through cameras, microphones and sensors, and with producing emotional signals in machines to make interaction feel more natural. Serious criticism distinguishes between detecting and producing the external pattern, on one side, and having an internal state, on the other.
Philosophical zombie is the thought experiment proposed by David Chalmers in 1996. An entity physically and behaviourally identical to a conscious human, but with no subjective experience inside. If such an entity is conceivable, consciousness doesn't reduce to behaviour or physiology. It's the sharpest image of what a system that mimics emotions without having them produces.
Qualia are the first-person felt qualities of conscious experience — the redness of red, the painfulness of pain, the bitterness of coffee. What is lost when a mental state is described only in physical or functional terms and the residue of "how it feels from the inside" remains.
Interoception is the perception of the body's internal state — heartbeat, breathing, visceral tension, temperature, pain. It's the sensory raw material on which many emotions are built, according to the theories that take the body seriously. A machine without that sensory network doesn't have that material available.
Referencias
William James (1884), What is an Emotion?, published in Mind 9, pages 188 to 205. Modern source of the question about the priority of body or mind in emotion.
Walter Cannon (1927), The James-Lange theory of emotions: A critical examination and an alternative theory, in The American Journal of Psychology 39, pages 106 to 124. The physiological critique that dismantles the strong version of James-Lange and opens the way to the Cannon-Bard theory.
Stanley Schachter and Jerome Singer (1962), Cognitive, social, and physiological determinants of emotional state, in Psychological Review 69, pages 379 to 399. The two-factor theory, arousal plus cognitive interpretation.
Antonio Damasio (1994), Descartes' Error: Emotion, Reason, and the Human Brain, Putnam, New York. Main source for the somatic marker hypothesis and for the functional role of emotion in decision.
Antoine Bechara and Antonio Damasio (2005), The Somatic Marker Hypothesis: A neural theory of economic decision, in Games and Economic Behavior 52, pages 336 to 372. Technical review of the hypothesis and of the Iowa Gambling Task as an operational test.
Joseph LeDoux (1996), The Emotional Brain: The Mysterious Underpinnings of Emotional Life, Simon & Schuster, New York. Source for the two routes of fear and for the distinction between defensive response and subjective experience.
Lisa Feldman Barrett (2017), How Emotions Are Made: The Secret Life of the Brain, Houghton Mifflin Harcourt, Boston. Source for the constructed theory of emotion and for the critique of Ekman's emotional universalism.
Paul Ekman (2003), Emotions Revealed: Recognizing Faces and Feelings to Improve Communication and Emotional Life, Times Books, New York. Popular exposition of the basic-universal-emotions thesis that the article discusses critically.
Rosalind Picard (1997), Affective Computing, MIT Press, Cambridge. Origin of the field and required reference for the claim of detecting and producing emotions in machines.
Kate Crawford (2021), Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence, Yale University Press, New Haven. The chapter on automatic emotion detection feeds the critique of affective computing as the industry sells it today.
David Chalmers (1995), Facing Up to the Problem of Consciousness, in Journal of Consciousness Studies 2(3), pages 200 to 219. Canonical formulation of the hard problem of consciousness that separates behaviour from subjective experience.
David Chalmers (1996), The Conscious Mind: In Search of a Fundamental Theory, Oxford University Press, New York. Development of the philosophical zombie argument used to distinguish simulation from experience.
Joseph Weizenbaum (1966), ELIZA — A Computer Program for the Study of Natural Language Communication between Man and Machine, in Communications of the ACM 9(1), pages 36 to 45. Reference for the first system that convincingly produced the external pattern of an emotional conversation with nothing inside.
Para profundizar
Plutchik, R. (1980). Emotion. A Psychoevolutionary Synthesis. Harper & Row. Origin of the "wheel of emotions", useful for understanding how old and how contested the claim to fix a universal emotional catalogue is.
Solomon, R. C. (1976). The Passions. Anchor. Philosophy of emotion predating contemporary neuroscience; retains value as a theoretical counterpoint to the biological substrate.
También te interesa
- The emotion-reason pairing. AI is only the rider
- Affective computing
- States of mind. No test measures the same person twice
No comments yet
No comments yet. Be the first.
Leave a comment