You think you're "seeing" other people — really you're rebuilding a whole brain for them, live, inside your own.
You're probably alone right now, yet a big chunk of your brain is dedicated to other people. Walk into a room and you read it almost instantly: who's angry, who wants to speak but didn't dare, whose "I'm fine" clearly isn't. You do it so smoothly you take it for granted. But this is one of the brain's most demanding, latest-maturing feats — building a model of someone else's mind inside your own, in real time, and constantly updating it. Better still, the model lets other people be wrong: you can hold the truth in one hand while tracking, in the other, the version in their head that differs from it. This issue is about that "social brain" — how it reads minds, how it empathizes, and one legend hyped to the sky and then dragged back to earth by science: mirror neurons.
Let's name the feat: theory of mind (also "mentalizing"). It's your ability to attribute invisible mental states — thoughts, wants, intentions, beliefs — to others, and use them to predict what they'll do next. When you say "he didn't pick up, he's probably busy," a whole act of mind-reading is hidden inside: you never saw his mind, yet you installed a "busy" state into it.
The fiercest part of the skill is that it can handle false beliefs. Psychology has a classic test, the false-belief task: a girl, Sally, puts a ball in a basket and leaves; while she's gone, Anne secretly moves it to a box. Now, the question: where will Sally look for the ball when she comes back?
You know the answer is the basket — because what you're tracking isn't where the ball actually is, but where Sally believes it is. That step looks trivial but is a huge cognitive leap: most children pass it reliably only around age 4, blurting "the box" before that (unable to keep "what I know" apart from "what she knows"). And it nests: I know you think I don't know — the everyday machinery of politeness, jokes, lies, and chess all runs on this layer-upon-layer recursive mind-reading.
AI has long wanted this, under the name machine theory of mind. DeepMind built a ToMnet: a network dedicated to observing other agents and guessing their beliefs and next move, able even to pass a machine version of the false-belief test. In multi-agent settings it's a necessity — to cooperate or compete you first need an "opponent model" (what does it probably want). And the last couple of years brought a lively fight: can large language models (ChatGPT and the like) read minds? Some found they answer many false-belief questions correctly; others tweak the wording slightly and watch performance crater. So whether the model is really mentalizing, or has simply seen oceans of human chat and learned "how mind-reading is talked about," nobody can yet call.
When you chew on "what is she thinking," a fixed cast of regions lights up together, collectively the mentalizing network. The leads: the medial prefrontal cortex (front-and-center-inward, handling "thinking about people and self"), the temporoparietal junction (TPJ, above and behind the ear, argued to specifically track others' beliefs), the precuneus/posterior cingulate at the back midline, and the temporal poles.
Here's the telling coincidence: this "mind-reading net" overlaps heavily with the network most active when you're doing nothing, just zoning out — the default mode network. In other words, the moment the brain goes idle, it defaults to thinking about people: replaying that conversation, decoding what someone meant, rehearsing what to say later. Social reasoning isn't a module you switch on occasionally; it's more like the brain's background hum, running whenever there's a free moment. Which is a sidewise proof that "humans are social animals" isn't a slogan — it's baked into the brain's default setting.
Say "understanding others" and many people reach for mirror neurons. This was an accidental find by Italian scientists in Parma in the 1990s, in monkey brains: certain motor neurons fired when the monkey reached for a peanut itself — and fired again when it merely watched someone else reach for one, sitting perfectly still. As if watching another's action, your brain quietly "ran through it" too. Humans turn out to have an analogous mirror system.
Then the story ran away with itself. A famous voice crowned them "the cells that built civilization": empathy runs on them, language began with them, autism is a "broken mirror." Stirring — but the evidence didn't keep up. Cooler-headed neuroscience dragged it back down: the mirror firing is real, but it does not equal mind-reading, nor empathy — what actually tracks others' beliefs is the mentalizing network from the last section, a different set of regions entirely. And the mirror property is likely, in large part, learned by association: whenever you perform an action you almost always see it too (you watch your own hand), so the brain ties "seeing" to "doing" — not necessarily an innate mind-reading organ. It probably handles "motor resonance, part of understanding actions, imitation" — a useful part, not the master key to social life.
"Watch once and run through it internally" resembles AI's imitation learning (learning from demonstration): an agent observes an expert's action sequence and maps it into its own motor system, getting most of the way without trial-and-error. The intriguing bit is that the same caution applies on both sides — a robot that only "mirrors" a human's motion trajectory often fails to grasp the intent and goal behind it, aping the surface and breaking in a new context. Which mirrors the mirror-neuron lesson exactly: copying an action ≠ understanding it. To truly get what someone (or an expert) is doing and why, you still need the mentalizing that builds a "goal/belief model." Mirroring alone isn't enough.
"Empathy" gets used too loosely; it actually bundles several different things with different circuits. At minimum, split it in two: cognitive empathy — you understand that he's hurting (via the mentalizing network, reasoning out his state); and affective empathy — you hurt along with him (watching another's pain lights up your own anterior insula and anterior cingulate — the very regions that process your pain). The two dissociate: some people grasp another's pain perfectly yet feel nothing; others tear up on sight but can't say what's actually wrong.
There's an even more practical cut. A line of work by Tania Singer found: affective empathy (hurting along) overused leads to burnout — that's empathic distress, why nurses and caregivers can collapse. But right beside it runs a different path: compassion — not "I hurt too" but "I care about you, I want to help," carrying warmth and the pull to act, riding a different (reward/affiliation-leaning) circuit. Crucially: compassion is trainable. Even brief loving-kindness meditation can shift people from "swamped by the pain" toward "warmly wanting to help" — a thread we'll pick back up in Phase C on meditation.
When you get a swarm of AI agents to cooperate, the useful part is usually not "feeling-with" but the cognitive empathy branch: give each agent a model of "what others want and how they'll move" and they can coordinate, yield, form alliances. It's practically the foundation of multi-agent cooperation. And the "affective empathy burns out, compassion is more sustainable" distinction quietly echoes an intuition in AI alignment: we don't actually want a system swamped and thrown off by every human pain in front of it; we want one that stably cares about and helps people — understanding them, caring, but not swept away by emotional contagion. Biology's answer (compassion ≠ empathic distress) happens to be exactly what the engineering wants.
"How do I ever get inside another person's mind" — the question is ancient, and several old traditions long ago moved in here: