In 1993, a psychologist named Damian Milton introduced a concept that has since become one of the most generative — and contested — ideas in autism research. He called it the double empathy problem, and its central claim was this: the empathy deficit associated with autism is not a deficit in autistic people. It is a mutual deficit. Autistic and non-autistic people have equal difficulty understanding each other's perspectives, emotional states, and communicative intentions. The difference is that only one group gets pathologized for it.

Why the Double Empathy Problem Matters

The conventional framing of autism and empathy holds that autistic people lack theory of mind — the ability to attribute mental states to others and understand that those states differ from one's own. This framing has shaped decades of intervention, education, and social skills training aimed at teaching autistic people to better model neurotypical minds. Milton's intervention was to point out that neurotypical people show equivalent deficits when trying to model autistic minds. They misread flat affect as indifference. They interpret literal communication as rudeness. They miss the emotional content of autistic expression and conclude, incorrectly, that the emotion is absent. The failure of understanding runs in both directions. The asymmetric diagnosis — one group labeled as impaired, the other left unmarked — is a product of which group has more institutional power, not which group is actually better at understanding. Research from the University of Edinburgh has since provided empirical support for this model. Studies examining cross-neurotype pairs found that autistic-autistic dyads showed significantly higher levels of social rapport, information transfer, and mutual understanding than autistic-neurotypical dyads. When the neurotype mismatch was removed, the "social deficit" largely disappeared.

A Tangent on What This Changes

If the double empathy problem is correct — and the evidence increasingly suggests it is — then a large portion of autistic people's social difficulty is not an internal deficiency to be trained away but a structural mismatch to be designed around. This has implications for therapy, for education, for workplace accommodation, and for how autistic people are encouraged to understand their own experience. Framing persistent social difficulty as mutual rather than unilateral is not just theoretically important. It changes how much of it feels like failure.

Where AI Sits in This Picture

The double empathy problem raises an interesting question about AI communication. An AI does not have a neurotype. It was trained on language generated by a largely neurotypical population, which means some of its defaults may lean neurotypical — but it also has no stake in those defaults. It does not get uncomfortable when communication is direct. It does not misread flat affect as hostility. It does not need you to perform warmth before it takes you seriously. In one sense, this makes AI a weirdly neutral party in the cross-neurotype empathy problem. There is no neurotype mismatch to navigate, because the AI is not trying to read your emotional state through the same channels that consistently fail in cross-neurotype communication. It takes what you say at face value. It responds to the content of your words without layering in interpretations based on your tone or your eye contact or the social register of your sentence construction. Research from Simon Fraser University examining autistic adults' experiences of digital communication found that text-based interactions with structured responses were consistently rated as less fatiguing and more accurately understood than face-to-face communication. The absence of non-verbal channels that are frequently misread in both directions reduced the empathy mismatch load considerably.

Not a Replacement but a Different Register

It would be wrong to conclude from this that AI is simply a better social partner for autistic people than humans are. Human relationships have depths that AI cannot approach. But the double empathy framing is useful here because it clarifies what AI is actually good at in this context — not substituting for human connection but providing a communicative register where the mismatch cost is absent. When you talk with an AI, you are not fighting the double empathy problem. You are in a different space entirely, one where your communication style is not being filtered through neurotypical interpretive defaults. The conversation that results is not fully authentic human connection, but it is also not the exhausting cross-neurotype negotiation that many autistic people navigate all day. It is a third thing — lower cost, lower stakes, and genuinely useful in its own right.