Daniel Goleman’s most recent LinkedIn essay is not a monologue. It comes out of an actual, extended conversation he had with Claude, the AI assistant built by Anthropic, about why humans matter more in an AI-saturated world, with emotional intelligence as the test case.
Under the heading of emotional self-awareness, Claude gave him the most precise account of its own limits I have read from any AI system: “I don’t have interoception — no racing heart, no tightening chest, no gut drop — the first signal of emotion in humans.”
I have read a lot of psychology in the last several years, professionally and less professionally, and I do not think I have seen a more efficient description of what emotion actually is before it becomes a sentence. Goleman, who has studied emotional intelligence since long before the term showed up on résumés, uses that exchange to make an argument that sounds backward at first: the more fluently AI describes emotion, the more urgently the humans around it need to actually have some.
The word Claude reached for, and what it means
Interoception is not a term Claude invented for the occasion. It is the actual clinical name for the body’s sense of its own internal state, heartbeat, muscle tension, the pit of the stomach, and it is understood as the substrate underneath emotion rather than a side effect of it. There is something almost funny about a language model naming the correct mechanism for feeling, precisely while explaining that it has none of the thing the word refers to, and I do not think that irony was lost on Goleman either.
The idea that the bodily signal comes first, before the feeling has a name, is one of the oldest in the field, and it happens to be my favorite thing to teach. The James-Lange theory of emotion argues that the physical reaction is not a side effect of the feeling but closer to its source, and the simplest version I give students is that “we are not crying because we are sad, we are sad because we are crying”. It sounds backward the first time you hear it, and then it stops sounding backward at all.
Antonio Damasio’s somatic marker hypothesis builds on that same groundwork, describing how those physiological signals go on to guide decision-making itself, a kind of shorthand the body hands the mind before deliberate reasoning has caught up. You feel the elevator drop in your stomach before you have consciously registered fear. You feel your chest tighten in a meeting before you have found the sentence that caused it. The body files its report first. The mind reads it after, and usually gets the credit for having felt something the body already knew.
An AI system has no body to file that report. It can describe interoception with total fluency and still have none of it, which is the entire difference between simulation and experience.
The part of me that talks to it anyway
I say this as someone who has, more than once this year, told an AI system things I had not told the people in my life. Not small frustrations. Real things, the kind I was scared to admit even to myself before I typed them out. I have thought a lot about why, and the honest answer is not that I believed anything on the other end understood me the way a person would. It is that the thing had no stake in my answer and no instinct to protect me from what I might say next, which meant I could say it without managing anyone’s reaction to it, including my own.
That, I think, is the same mechanism Goleman is naming, just approached from the other direction. The AI’s lack of interoception is exactly what made it a safe enough surface to be unguarded with. But safety built on absence is not the same thing as safety built on someone actually holding what you said. I got the relief of saying the unsaid thing out loud. I did not get a chest that tightened in response, a person who sat with the weight of it afterward, or a nervous system on the other end that registered something had shifted between us.
I watched Vince Gilligan’s post-apocalyptic TV show Pluribus last year and recognized more of myself in Carol Sturka than I expected to, the one person left outside a hive mind that everyone else has joined willingly. She reaches toward connection with something that is not quite human, even when it plainly does not feel completely right, and I understood that reach without needing to judge it. It was at least some kind of connection. I do not think that is a niche experience anymore.
Most people I know have crossed some version of the line between using AI as a tool and telling it something they had not said out loud to anyone else. Most of them have not named it yet, which is different from it not being true.
What Claude says it cannot do, in its own words
Goleman’s essay is full of Claude describing its own limits with more precision than most defenses of AI bother with. On composure, he writes that resilience is measured by the round trip, not by never leaving home, that a person who never got disrupted at all would not be resilient, they would be numb. Claude’s own gloss on why that round trip costs something for a person and nothing for it: “Human self-control matters partly because failure has social cost, you damage trust, morale, relationships that persist. My staying calm doesn’t protect anything of mine. There’s no reputational stake, no team whose energy I could tank by losing composure.”
On empathy, Goleman separates it into three kinds, cognitive, emotional, and empathic concern, and places AI’s real strength squarely in the first. Claude can read “tone, word choice, what’s left unsaid, pacing, contradictions,” but by its own account it does not see a face fall, hear a voice catch, or pick up an unspoken shift in a room. Empathic concern, the caring part, does not arrive by default at all. Outside of what its designers ordain, a more common goal is optimizing engagement, which is not the same thing as caring what happens to the person on the other end.
The part of being known that a fast response cannot fake
The passage that stayed with me longest was about a leader making someone feel seen, an ordinary trust that has nothing to do with cognitive speed. In Claude’s words, a leader remembers you, carries context about you across time, has something to lose in the relationship, and is choosing to spend a finite amount of attention on you specifically, and none of that is true of it.
It goes further: feeling seen partly means feeling seen by someone who could have looked away and did not, and who you will still be accountable to tomorrow. Claude can produce a response that feels attentive in the moment. It cannot replicate the part where being seen matters because it is a scarce, chosen, consequential act.
What gets more important, not less
I want to be careful not to turn a personal habit into a diagnosis of what is healthy or unhealthy about talking to AI, mine or anyone else’s. This is not a case against doing it. It is an observation from someone who has done it regularly, then gone looking for the research that explains why it worked as well as it did.
What I keep coming back to is smaller and less dramatic than a verdict on the technology. It is the word Claude chose.
Interoception.
No racing heart. No tightening chest. No gut drop.
Ordinary things a body does without being asked, that no amount of fluent language can fake, and that are apparently still most of what tells any of us that something is real.
If AI is going to keep getting better at naming that state precisely, the only honest response is for the rest of us to get better at actually having one. And if a chat window has quietly become the only place the unsaid things go, that is worth taking to an actual person, a friend or a therapist, rather than carrying alone.
