Does the double empathy problem hold up?

Because it named something autistic people had been saying for years. Jim Sinclair, an autistic writer, had argued as early as 1993 that the empathy gap runs both ways, and in 2010 described how “being autistic in shared autistic space may be easier than being autistic in non-autistic space”. Autistic communities had long talked about the relief of being among their own. What was missing was a research frame that took the observation seriously instead of treating it as a symptom.

The double empathy problem gave it one. It also arrived as the theory it challenged was losing ground. The claim that autistic people lack a “theory of mind” had dominated since the 1980s, and by the 2010s researchers were documenting its empirical failures. Milton’s reframing offered an alternative that fit both the community’s experience and the emerging data, and it came from inside the community. That combination is rare, and it travelled.

Milton’s paper, “On the ontological status of autism: the ‘double empathy problem’”, appeared in the journal Disability & Society in 2012, in its “current issues” section. It is five pages long. It is not an experiment. It’s an argument, drawn from sociology and philosophy, that opens and closes with Plato and cites Garfinkel, Becker, and Said rather than any psychology lab. Its author, then at the University of Birmingham, writes as an autistic self-advocate, and says so in the first paragraph.

That matters for reading everything that follows. The founding document was a proposal, not a finding. Its author said so plainly ten years later: the concept was “initially derived through a combination of personal experience, anecdotal accounts and limited qualitative data”. The evidence came afterwards, much of it from researchers who hadn’t read the paper.

The paper’s own definition is worth having in full. The double empathy problem is “a disjuncture in reciprocity between two differently disposed social actors which becomes more marked the wider the disjuncture in dispositional perceptions of the lifeworld”. In plainer words: the further apart two people’s ways of experiencing the world, the more likely they are to misread each other, and the misreading belongs to both of them. “It is a ‘double problem’ as both people experience it, and so it is not a singular problem located in any one person.”

Three claims follow from that, and the paper makes each of them.

First, that autism has been defined as a “social deficit” inside one person, and that this is a choice rather than an observation. “Differences in neurology may well produce differences in sociality, but not a ‘social deficit’.” Second, that the deficit had been measured from one side only: non-autistic “empathy”, applied to an autistic person, “is often wildly inaccurate in its measure”, and the “lack of theory of mind myth persists” anyway. Third, and least noticed, an asymmetry. Autistic people, the paper argues, have often gained “a greater level of insight into NT society and mores than vice versa”, because they have to live in it, while a non-autistic person “has no pertinent personal requirement to understand the mind of the autistic person”.

What the paper does not quite say, though it’s implied by the definition, is that autistic people will understand each other more easily. That prediction, the one most of the later experiments test, was drawn out afterwards. Milton framed it in 2022 as “greater potential”, not a guarantee.

The paper was also a direct critique of the theory-of-mind account, of behavioural intervention aimed at making autistic children “fit in”, and of what it calls an autism industry that silences autistic voices. Those arguments travelled less far than the definition did, and they’re the ones the field is now circling back to.

It reframed a question. Before 2012, research on social interaction in autism asked what was wrong with the autistic person’s social cognition. After it, a small number of researchers began asking a second question: how do non-autistic people fare at reading autistic people?

That reframing is the paper’s whole contribution and it’s a large one. It didn’t measure anything, and the author has never claimed it did. Several of the studies below were designed without reference to it. Milton noted in 2022 that many researchers whose findings support the concept “were initially unaware of the concept”, and only later framed their work in its terms.

The first tests came from the non-autistic side, and they mostly used the same trick: show non-autistic people real autistic behaviour without telling them it’s autistic, and see what happens.

Sheppard and colleagues in Nottingham, in 2016, filmed people reacting to one of four greetings and asked non-autistic viewers to guess which greeting they’d received. Viewers were significantly worse at reading autistic people’s reactions, without knowing anyone was autistic. A separate group rated the autistic people as equally expressive. The signal wasn’t weaker. It was different.

Edey and colleagues, the same year, had autistic and non-autistic adults animate simple shapes to act out emotions such as coaxing and mocking, then asked non-autistic viewers to name the emotion. Viewers were much less accurate with animations made by autistic people. The authors titled the paper “Interaction takes two”.

Then Sasson and colleagues in 2017 pulled three independent studies into one paper, and it became the most-cited evidence in the field. Non-autistic observers watched or heard a few seconds of real behaviour from autistic and non-autistic people, again without being told anyone was autistic. Autistic people were rated as more awkward, less likeable, and less attractive, and observers were less willing to sit near, talk to, or befriend them. The effect appeared within seconds, held for children as well as adults, held for a single still photograph, didn’t change with repeated exposure, and vanished when observers read a transcript of the words alone. Style, not substance. Ratings of intelligence and trustworthiness didn’t differ.

Heasman and Gillespie, in 2018, looked at families. Autistic adults and their relatives each predicted how the other would describe them. Both sides expected to be misunderstood. In practice, neither side actually misjudged the other much. Perspective-taking, the authors wrote, “is two-sided”.

The 2016 and 2017 studies showed one direction of the problem. The double empathy claim also predicts something about autistic people together: if the difficulty is a mismatch, it should shrink or disappear when the mismatch does. The clinical account predicts the opposite. Two autistic people should struggle most of all.

Crompton and colleagues in Edinburgh tested that head-on in 2020. Seventy-two adults were sorted into nine chains of eight people: three chains all autistic, three all non-autistic, three alternating. A researcher told the first person a thirty-detail story about a bear. That person told the next, and so on down the chain. The all-autistic chains kept as much of the story as the all-non-autistic chains. The mixed chains lost detail significantly faster. Afterwards, people in the mixed chains reported lower rapport with their neighbours. The groups were matched on age, education, and IQ, and the analysis was pre-registered.

The same team interviewed twelve autistic adults about time spent with autistic and non-autistic friends and family. Three themes: they felt better understood by other autistic people and found non-autistic company effortful and exhausting; they felt like a minority pressed to conform; and among autistic people they felt they belonged and could drop the mask. One participant, on discovering autistic friends: “I never realised everybody felt as happy as I do when I am around autistic people.”

In the United States, Morrison and colleagues paired 125 autistic and non-autistic adults for five-minute conversations with a stranger, in all three combinations. Autistic people were rated more awkward and less warm by both autistic and non-autistic partners. But only non-autistic people preferred future interaction with non-autistic partners. Autistic people leaned towards other autistic people and disclosed more to them. The authors described social difficulty in autism as “relational rather than an individual impairment”. Our reading of this study is from its abstract; we have not read the full paper.

Two theoretical papers then gathered the strands. Mitchell, Sheppard, and Cassidy, in 2021, proposed a developmental loop: non-autistic people misread autistic behaviour, judge it unfavourably, and act unwelcomingly; autistic people notice, withdraw or camouflage, and become harder to read; and round it goes, with camouflaging and thwarted belonging as the route to poor mental health. Davis and Crompton, the same year, connected the evidence to a second framework, the dialectical misattunement hypothesis, and argued the field needs developmental studies and measures of social skill that aren’t normed on non-autistic behaviour. Milton, Gurbuz, and López marked the tenth anniversary in 2022 with a short editorial noting that research had moved from anecdote to “numerous research projects” across psychology, neuroscience, linguistics, and design.

The evidence refined the claim as much as it supported it. Four refinements matter.

Autistic raters aren’t kinder. The popular version of the idea says autistic people understand each other and non-autistic people don’t understand them. The studies say something narrower. When autistic observers rated autistic people, in DeBrabander and colleagues’ 2019 study and in Morrison’s, they rated them less favourably too, much as non-autistic observers did. The difference was that autistic raters weren’t put off. They still wanted to interact. So the evidence supports easier exchange, higher rapport, and preference, not a mirror-image ability to read minds.

It may not be about autism at all. Alkhaldi, Sheppard, and Mitchell, in 2019, showed the same videos from the Sheppard study to new raters. People who were hard to read were rated unfavourably, and that link held regardless of diagnosis. Being autistic made you more likely to be hard to read, but the mechanism looked general: we like people we can interpret. Mitchell and colleagues now describe the pathway as transdiagnostic.

The direct test is one experiment. The diffusion chain study is powerful, pre-registered, and widely cited, and its authors list its limits in full. Seventy-two people. Participants knew whether their chain was mixed. Three of the autistic participants were self-identified. Most were women with typical or high IQ, diagnosed as adults. The rapport finding rested on modest statistical evidence for the autistic-versus-mixed comparison. The authors call for replication, and as of the papers we read, the field is still waiting for a large one.

Only one study tested whether autistic people read autistic people better, and it used animated shapes rather than real behaviour. Mitchell and colleagues point this out as the largest gap. The claim that autistic people are better at reading each other’s inner states, as opposed to getting on better, has been assumed more than measured.

The language. Sasson’s 2017 paper, which did more than any other to establish the non-autistic side of the problem, is written throughout in “impairments”, “deficits”, and “typically developing”. Its own conclusion cuts against its vocabulary: the authors suggest that “educating others to be more aware and accepting” may be more sensible than training autistic people to manage impressions. Later papers from the same group use “non-autistic”. Reading the early evidence means reading past the frame it came in.

The one-sidedness of the critique. The theory-of-mind account that Milton was arguing against has been substantially dismantled on its own terms, in work such as Gernsbacher and Yergeau’s 2019 review of its empirical failures. The double empathy problem no longer needs to define itself against a strong opponent, and the more recent papers don’t.

And the popular retelling. “Autistic people understand each other perfectly” is not what the paper claimed, not what the evidence shows, and not what Milton says. Two of the twelve people in Crompton’s interviews found other autistic people hard work, one because of bluntness and one because of unpredictability. The research supports greater potential for understanding, which is what was proposed.

The bidirectional claim is well supported. That non-autistic people misread and misjudge autistic people, quickly and on style rather than content, has been shown by several independent groups using different methods, and it replicates. That autistic people exchange information and build rapport with each other at least as well as non-autistic people do with each other has been shown once, carefully, and is consistent with interview and conversation studies.

What follows from it is where the researchers themselves now put their effort. Mitchell and colleagues’ developmental model is explicit that its links to camouflaging, belonging, and suicidality are hypotheses to be tested, not findings. Davis and Crompton’s call for autism-specific measures and longitudinal designs is a description of work not yet done. Milton’s 2022 editorial asks for research that takes power, intersectionality, and “participatory sense-making” seriously, and for autistic people to be involved at every stage. The field has a supported premise and an agenda.

Mechanism. Nobody knows why mixed pairs lose information or rapport. Different social cues, different processing, anxiety about the other group, non-autistic people simplifying “for the autistic person’s benefit”: the diffusion-chain authors raise all of these and test none.

Whether familiarity fixes it. Repeated exposure to photographs didn’t soften first impressions. Real acquaintance might. Sasson’s group found that telling raters someone is autistic improves impressions, but Heasman and Gillespie found that non-autistic people overestimate how helpful they then are. No study has followed cross-neurotype relationships over time.

Development. Every study here is a snapshot of adults, or a snapshot of children rated by strangers. Whether the gap widens across a childhood, as the transactional model predicts, hasn’t been observed.

Who it applies to. Almost every participant in this literature is a speaking adult with typical or high IQ, in the UK or the US, and most are women diagnosed as adults. Autistic people with intellectual disability, non-speaking autistic people, children, and people outside two countries are missing from nearly all of it. Milton’s editorial names the next question directly: how do two people of different ethnicities and different neurotypes experience an interaction? Nobody has looked.

Mental health. The link from being misread to camouflaging to thwarted belonging to suicidality is the most consequential claim in this literature, and it’s a model, drawn from correlational pieces, awaiting a test.

This is an idea that came from autistic people and was first published by an autistic scholar. The strongest experimental group working on it includes an autistic co-author with a background in advocacy, and its interview schedule was reviewed by autistic people. The 2022 anniversary editorial makes participation a condition, not a courtesy.

The ethical stakes are practical. If the difficulty is inside the autistic person, social skills training and impression management are the rational response, and that’s what autistic children have received for decades. If the difficulty sits between people, half of that effort is aimed at the wrong side. Mitchell and colleagues go further and argue that interventions built on the deficit view may themselves produce the thwarted belonging they link to suicide risk. That’s a hypothesis. It’s also the reason this question isn’t academic.

Language shifted across the decade covered here, from “typically developing” and “impairment” to “non-autistic” and “difference”, and from “people with autism” to “autistic people”, which most autistic adults prefer. This guide uses the later terms and quotes the earlier ones where they’re the authors’ own.