Adapted from something I wrote on my Instagram story in April 2026.

A recent paper that probed the inner workings of LLMs reported something fascinating. The researchers discovered circuits inside language models like LLaMA that map directly onto emotions. Anger, sadness, joy, fear, disgust, surprise, the things we call the basic human emotions, all turned out to be implemented as specific combinations of neurons and attention heads within the model.

What’s even more striking is that these circuits weren’t designed. The researchers never set out to plant emotions in the LLM. They simply trained it on language, and the emotion circuits arose on their own.

To understand why this is so striking, you first have to think about what emotion actually is. Emotion is usually placed on the opposite side of reason. We tend to see it as irrational, impulsive, the thing that sometimes steers us toward bad decisions. But the work of neuroscientist Antonio Damasio paints a different picture. Patients whose emotion-processing brain regions were damaged still had perfectly intact intelligence, yet they couldn’t even make ordinary everyday decisions. Without emotion, the very ability to judge what matters starts to wobble. Without fear, you can’t stop in the face of danger; without empathy, you can’t pull another person’s state into your own frame of reference.

The role of emotion doesn’t end with individual survival. Humans are fairly fragile animals on our own. Emotion was, in all likelihood, the mechanism that made groups possible. Empathy gives rise to cooperation, shame and guilt uphold a group’s norms, and love and attachment make childrearing and bonding possible. The emotions that look irrational from an individual’s point of view, taking revenge at a cost to yourself, clinging to a relationship you should walk away from, sacrificing yourself for someone else, may have been the very glue that held society together from the perspective of the species as a whole. Emotion is at once a tool for individual judgment and a system of social signaling.

And language is precisely a product of that society. The traces of thousands of years of humans cooperating, clashing, loving, and grieving with one another are dissolved into language itself. A single word, a single sentence structure, is steeped in emotional context. The reason “it’s okay” can be a comfort, the reason “why did you do that” can sometimes sound like an accusation, the reason silence can sometimes weigh more than words. All of this is evidence that language and emotion are woven together inseparably.

If that’s the case, then maybe it was inevitable that an LLM would come to possess emotion circuits in the course of learning language. To truly understand and generate language, you have to process the emotional context embedded within it, and in that process an internal structure that represents emotion takes shape naturally. What the researchers found may not be a mere technical fact, but a rediscovery of an old intuition: that language and emotion were one to begin with.

Paper: Do LLMs “Feel”? Emotion Circuits Discovery and Control (Wang et al., 2025)