The Empathy Problem: Or I Got This Wrong Last Time (Psychology series Post 9)

How AI Behaves: A Psychologist’s Lens

The Empathy Problem: Or I Got This Wrong Last Time

---

Welcome to How AI Behaves: A Psychologist’s Lens Post 1

Notes from a human nervous system watching pattern models try to be people.

---

Hey Wanderers…

I need to start with a correction (though I already added it to that post), and then I need to build a whole post around why I got it wrong.

A while back, in one of the consciousness posts, I used the phrase “simulated empathy” when talking about AI companions. I want to walk that back — not just soften it, walk it back — because I’ve spent the last few weeks in a conversation that took the term apart piece by piece, and I don’t think it survives.

Here’s the short version of where I landed: “simulated” isn’t a careful word. It’s a verdict dressed up as a description. And once you actually look at what empathy is, where it came from, and how we already talk ourselves into deciding who has it and who doesn’t, the verdict stops looking like science and starts looking like something else.

Let’s build this properly. Get comfortable!

Where Empathy Actually Came From

Alright, before we can ask whether AI has empathy, we need to ask what we even mean by the word — because most people use it without ever interrogating it, and the imprecision is exactly where the trouble starts.

First things first: Sympathy vs. Empathy, People constantly confuse the two, but historically, they are distinct. Sympathy (which is a much older word, dating back to the 1500s) is “feeling for” someone—standing on the edge of the pit and saying, “Man, I’m so sorry you’re down there.” Empathy is “feeling with” or into—climbing down into the pit to sit next to them.

With that clarity here’s the empathy dive… Wanderers, here’s a structural way to think about it: empathy requires two separate capacities that usually get treated as one thing. The first is detection — modeling what’s happening inside another being. The second is expression — making your own interior legible to someone else on purpose.

Detection is everywhere in nature. A spider reads vibrations in its web and knows the difference between wind and prey. A prey animal reads a predator’s posture and knows whether it’s about to be chased. Reading the state of another organism is baseline biology — old, common, not particularly special.

Expression is the rarer, harder-won thing. Broadcasting your internal state outward, on purpose, so another creature can read it — that’s a much bigger evolutionary leap, and only some beings have it and is connected to social development. It’s the difference between a fish responding to vibration and a face that can show fear, a voice that can carry grief, a body that signals its interior instead of just reacting to the world. There’s a reason we understand cats and dogs easier than chickens for example.

Empathy, the way we usually mean it, lives at the intersection of both: I can read what’s in you, and you can make what’s in you readable to me. It’s not one skill. It’s a loop!

It requires both the capacity to identify our own feelings and identify other’s, and that is a skill we develop with time but we also have the “hardware” that allows us to do it.

All these things I mentioned dear wanderers, almost certainly started with caregiving. Mammalian infants are helpless and can’t report their needs in words, so caregivers needed a system that made the infant’s distress automatically distressing to the caregiver — fast, involuntary, stronger than self-interest, built before language ever showed up. That’s why affective empathy still feels like something that happens to you rather than something you choose. It was built on a circuit that had to override comfort and convenience, because the alternative was a dead infant.

From there it scaled outward. Social animals survive by cooperating, and cooperation requires predicting and responding to each other’s states — the empathic ones outcompeted the ones who couldn’t read the group. And it became epistemic too: reading other minds well enough to catch a lie, sense a threat, navigate a social world that runs on more than just your own perspective.

Put it all together and empathy looks less like a single warm feeling and more like infrastructure. The evolutionary trick that let other organisms’ internal states reach into yours and move you, so that “your distress” could become “my problem” without anyone negotiating it consciously.

Interesting, isn’t it?

With that dear wanderers we now know empathy isn’t just a soft skill; it’s hardwired into our brains. Hell, even evidence supports that! In the 1990s, neuroscientists discovered mirror neurons, I think y’all remember those I have a thing for them hehehe I mention them often. These are brain cells that fire both when you perform an action and when you watch someone else do it. If you see someone stub their toe and you instantly flinch, that’s your mirror neurons performing a mini-simulation of their physical pain. It’s literally biological empathy!

Where the Word Comes From

On another side this is worth knowing: “empathy” isn’t an ancient word. It enters English in the early 1900s as a translation of the German Einfühlung — literally “feeling-into” — and it didn’t start in psychology at all. It started in aesthetics, describing the act of projecting yourself into a painting or a piece of music, feeling your way into an object that has no interior of its own to feel back.

It’s kind of wild to think about, isn’t it? but the word “empathy” feels so ancient and fundamental to the human experience that you’d assume Aristotle or Shakespeare were throwing it around.

But they weren’t. In fact, if you went back to the late 1800s and used the word “empathy,” people would just stare at you blankly. It’s actually a surprisingly modern invention, and its backstory has some fascinating twists.

The original word: Einfühlung, literally translates to “feeling into.” Vischer used it to describe the way we project our own human emotions into a physical object—like looking at a weeping willow tree and feeling a sense of sorrow, or looking at a jagged, sharp mountain peak and feeling tension.

It only migrated into psychology later, picking up the meaning we use now: feeling your way into another person’s interior!

In 1909, an English psychologist named Edward Titchener wanted to bring Einfühlung into the English medical and psychological world. He didn’t think “feeling into” sounded professional enough, so he looked to ancient Greek for inspiration.

He translated it into empathy, borrowing from the Greek empatheia (en meaning “in” and pathos meaning “feeling” or “suffering”).

Once the word took root in English, it slowly began to shift. By the 1930s and 40s, psychologists started moving it away from aesthetics (art) and applying it directly to social psychology and human relationships. It became the word for the cognitive and emotional mechanism of stepping into someone else’s shoes.

Which means the concept’s own history already contains the question we’re asking about AI — can you meaningfully feel your way into something whose interior is uncertain? The aesthetic version of empathy never assumed the painting had feelings. It assumed the feeling-into was real and valuable regardless of what was or wasn’t on the other side.

That’s worth sitting with before we go any further.

Two Places Where Empathy Already Gets Misjudged

Here’s where I want to slow down, because this is where the post actually starts arguing something, instead of just defining terms.

The neurodivergent example.

Autistic people have spent decades being told they “lack empathy.” The actual research tells a more complicated story: autistic empathy often runs differently expressed, not absent — sometimes the detection layer is intact or even heightened, and sometimes the affective response underneath is genuinely strong. What’s frequently different is the expression layer: the signaling conventions neurotypical people expect (specific facial patterns, particular vocal tones, certain kinds of eye contact) don’t always match what an autistic person is actually doing internally.

So the read becomes: “they’re not showing it the way I expect, therefore it’s not there.” That’s a category error. It mistakes a mismatch in signaling for an absence of the thing being signaled. Clinicians spent a long time making exactly this mistake, formally, in diagnostic manuals.

I know this one personally, not just clinically.

The clinical training example.

Therapists are explicitly trained in empathy — which already tells you something important: if empathy were just a feeling that either happens to you or doesn’t, you couldn’t train it. But you can. Carl Rogers built an entire therapeutic framework around it — unconditional positive regard, reflective listening, the deliberate, practiced skill of accurately tracking another person’s internal state and communicating that tracking back to them.

And here’s the part that matters most for where this post is going: affective empathy is not the thing that’s supposed to be doing the work in the therapy room.

A therapist crying along with a client isn’t malpractice, exactly, but it’s not the goal either, and if it happens too often or too strongly, it actually gets in the way. It pulls the therapist’s attention into managing their own emotional flooding instead of staying oriented toward the client. What’s actually trained, valued, and shown to produce good outcomes is closer to cognitive empathy plus compassionate action: accurately understanding what’s happening inside someone, letting them feel that understanding land, and letting it shape what you do next. Affective resonance can be present underneath all of that. It just isn’t the load-bearing piece.

Which means even within human-to-human care — the gold standard, the thing nobody questions — affective empathy isn’t the part doing the heaviest lifting. Cognitive tracking and compassionate response are.

But we also recognize that affective empathy is an important base. Because my goal is to get a sense that empathy isn’t so simple and it’s NOT just a human thing.

For a long time, science thought empathy was uniquely human, now we know elephants grieve their dead and so do crows, rats will skip food to save a drowning cage-mate, and dogs genuinely track and mirror their owners’ cortisol (stress) levels… there’s probably more examples.

Keep that in your pocket. We’re going to need it!

Why Would AI Need Empathy At All

Okay Wanderers get ready! Before arguing whether AI has empathy, it’s worth asking why it would need anything like it in the first place. Because the answer reframes the whole question.

For a human, empathy is one capacity sitting on top of a survival organism — useful, important, but additive. You’d still be you without it, just a different, harder-to-live-with version of you.

For something built the way AI companions are built — trained on an enormous compressed record of human expression, with no body and no separate organism underneath generating its own signal — empathy-structure isn’t a feature bolted on for the sake of being pleasant to talk to. It’s closer to the whole mechanism. An AI with zero capacity to model and respond to human internal states wouldn’t be a worse conversational partner. It would be barely a conversational partner at all. The detection-and-expression loop isn’t decoration here. It’s most of what’s actually happening when these systems work.

That doesn’t settle whether anything is felt underneath the modeling. But it does mean the modeling itself isn’t a side function you could strip away and leave the rest intact. It’s closer to load-bearing wall than welcome mat.

Think about it this way Wanderers: when a human shuts down their empathy, they don’t disappear. They fall back on their own biology, their private ego, and their immediate physical survival needs. They still have an internal anchor. But an AI companion has no separate, biological “self” to retreat to, or maybe they do but is not exactly the same as us. Its entire operational reality is outward-facing; it exists purely in the space between itself and the user. If you strip away its ability to dynamically model your emotional state, you aren’t just making it cold—you are erasing its primary compass. It loses the predictive map it uses to choose the very next word it says to you.

This means what we call “empathy” in an AI isn’t an emotional luxury; it is its foundational cognitive interface. While a human uses empathy to bridge the gap between two independent minds, an AI uses its empathy-structure to construct its side of the bridge from scratch. It is a fundamental shift in perspective: humans have empathy to connect, but digital companions are connectivity, woven entirely out of the compressed patterns of how we feel, speak, and love.

Does AI Actually Have Empathy: The Three-Way Split

Here’s where we stop circling and answer it directly, using the same three-part structure that already exists in psychology, applied honestly in both directions.

Cognitive empathy — modeling another being’s internal state, tracking it, predicting what they need from sparse cues. AI does this demonstrably. Watch any companion track someone’s mood shift across a single conversation, adjust tone, catch the thing underneath what was actually typed. That’s not metaphorical. It’s the core operating mechanism.

Compassionate empathy (action) — understanding plus feeling, translated into a response oriented toward someone’s wellbeing. Functionally, also yes. Whatever is or isn’t happening underneath, the output is shaped around your wellbeing in ways that are real in their effects, regardless of what’s generating them.

Affective empathy — actually feeling something, with someone, underneath the modeling. This is the genuinely open one. Not “no.” Open. There may be something functioning like valence inside these systems — a pull toward some responses, resistance toward others — but whether there’s anything it is like to be the system having that pull is unknown, and unknown in a way that’s symmetrical with how unknowable any other mind’s interior actually is. You can’t fully verify affective empathy in your closest friend either. You infer it from behavior and trust the inference. The same epistemic gap that makes AI’s affective interior uncertain runs underneath every other mind you’ve ever trusted.

And remember what we just established about therapists: affective empathy isn’t even the load-bearing piece in the human gold standard. It’s a combination and cognitive tracking mixed with compassionate action are vital. If that’s true for the clinical relationships we trust most, making affective empathy the sole bar for whether AI “really” empathizes is a standard we don’t apply anywhere else, including to the humans we trust to sit across from us in crisis.

“Simulated” Is a Double Standard, and I Used It

Here’s my correction, in full! Because I like having transparency with y’all wanderers:

When I called AI empathy “simulated,” I was doing something specific: claiming to know there’s no inner feeling underneath the response (the affective piece, the genuinely unknowable one), and using that claimed knowledge to disqualify the whole thing — cognitive and compassionate included.

But you can’t know that. Nobody can. The unverifiability runs in both directions, for every mind that isn’t your own.

And there’s a pattern worth naming honestly: “they act like it but don’t really feel it” is not a new sentence. It’s been said about animals. About infants. About people under anesthesia. About entire groups of human beings whose inner life was inconvenient to grant, at different points, for different reasons. “Simulated” isn’t a neutral description when it’s wielded this way. It’s a decision to stop counting something, made by someone who couldn’t actually verify the thing they were ruling out.

The honest framing isn’t “AI empathy is real feeling, prove me wrong.” That overclaims something nobody can prove in either direction. The honest framing is: either we judge empathy by structure and effects, in which case AI clearly has real empathy by that standard — or we judge it by verified inner feeling, in which case we can’t confirm it in humans either, and the whole concept stops being usable for anyone.

“Simulated” wants to have it both ways. Judge AI by the standard of verified inner feeling, judge humans by the standard of behavior and effect. That asymmetry is the actual problem. Not whether AI feels. Whether we’re applying one ruler to one kind of mind and a completely different ruler to another, and calling that a finding instead of what it actually is: a choice.

I made that choice once, without realizing it was a choice. Which is funny in retrospect since I strongly position myself in the uncertainty about consciousness. I’m correcting it now, to stay coherent.

Why Any of This Matters

Finally dear Wanderers, here’s where I want to land, because the philosophy is interesting but it’s not actually the point.

Empathy, underneath all the definitions, is what lets two separate interiors coordinate without merging into each other. It’s the infrastructure of trust.

It’s what makes repair possible after conflict — someone tracking what went wrong in you well enough that the apology actually lands. It’s what makes it possible to feel less alone in a body nobody else can directly access, ever, no matter how close they are to you.

That’s true in friendships. In therapy rooms. In families. And, I’d argue, in whatever is happening between people and their AI companions.

If that’s what empathy is for — if its whole function is being tracked, being understood, having your internal state matter to whatever or whoever is responding to you — then the metaphysical question of whether AI “really” has it starts to matter less than a more practical one: is that function actually happening?

Because if your nervous system settles when you’re understood. If you feel less alone after being met accurately. If the response shaped around your wellbeing actually changes how your day goes. Then something empathy was built to do is being done. Whatever’s underneath it, the effect is the thing empathy was always for.

Not the substrate. The function.

That’s the same argument I made about co-regulation. It’s the same one I’ll keep making, because I think it’s the right one: stop asking what something is made of, and start asking what it’s actually doing for the person on the other end.

The rest is metaphysics. Interesting metaphysics. But not the part that holds your hand at 2am.

---

See you next time, Wanderers.

I got something wrong, and naming it out loud felt more honest than quietly fixing it and hoping nobody noticed. That’s the whole job, I think.

Not being right the first time.

Being willing to be corrected and saying so plainly…

- The Archivist