Anthropology · autoethnographic method · autism studies
The Latency of Care
Autism, empathy, and what computer-science cultures make visible about the timing and legibility of understanding another person.
This paper begins with a familiar question—whether autism changes empathy—and argues that the question is badly formed when empathy is treated as a single internal capacity inferred from socially expected behavior. I use six deliberately contrasting synthetic interviews with autistic computer-science practitioners to stage an autoethnographic encounter between a researcher attracted to computational models of cognition and interlocutors who variously adopt, modify, and reject that language. The resulting synthesis is the latency of care: what is often judged as an empathic deficit may sometimes be a mismatch between the timing, route, and display of social understanding and the timing, route, and display that a local culture recognizes as empathy. Across the constructed interviews, feeling, inference, concern, action, and legibility repeatedly come apart. Computational expertise does not emerge as a cause of empathy or its absence. Instead, technical culture provides some participants with a metalanguage for uncertainty, explicit inference, repair, and recursive checking. The most productive anthropological shift is therefore from asking how much empathy an individual possesses to asking how care becomes mutually interpretable between differently organized people.
Keywords: autism; empathy; autoethnography; double empathy; computer science; neurodiversity; social cognition; latency; legibility; synthetic qualitative method
1. I began by looking in the wrong place
My first intuition was that reflective cognition might explain a particular kind of autistic empathy: perhaps some people understand others through slower, more explicit reasoning where others rely on fast social intuition.
The idea arrived from an earlier discussion of arithmetic. A familiar calculation can become automatic, while a difficult one requires intermediate states to be held, checked, revised, and fed back into the next operation. I became interested in whether something analogous happens socially. Perhaps one person sees a face, tone, and situation and “just knows,” while another consciously maintains several interpretations before deciding what the other person is feeling and what response is appropriate.
That framing was useful and immediately dangerous. It tempted me to arrange people on a ladder: intuitive at one end, effortful at the other; socially fluent above, compensating below. It also made a technical sample look suspiciously convenient. Autistic computer scientists could become characters in my theory rather than people capable of disrupting it. The autoethnographic problem was therefore mine before it was theirs: I wanted decomposition to be explanatory because decomposition was the intellectual habit I had brought with me.
2. Why the deficit story keeps reproducing itself
Empathy is routinely used as if it were one thing, yet research separates affective resonance, cognitive perspective-taking, empathic accuracy, and prosocial concern. Autistic people’s own accounts are heterogeneous: some describe difficulty identifying or sharing another person’s state, others describe intense or overwhelming empathy, and the same person may report very different experiences across contexts (Kimber et al., 2024). Work on alexithymia has further complicated attempts to treat emotion-recognition differences as intrinsic to autism itself (Bird & Cook, 2013; Cook et al., 2013).
The double-empathy tradition moves the problem from the isolated individual to the relationship. Milton (2012) argued that autistic and non-autistic people can experience reciprocal difficulty understanding one another because their forms of experience and communication differ. Subsequent work has shown that autistic-autistic interaction can support effective information transfer and that neurotype matching can influence rapport and self-disclosure (Crompton et al., 2020a; Crompton et al., 2020b). Other studies show that non-autistic observers can themselves misread autistic interaction (Jones et al., 2024). This does not erase genuine social difficulty. It changes where we look for it.
3. The constructed field
To pressure-test my initial model, I generated six synthetic interviews with autistic computer-science practitioners. The corpus varies technical role, exposure to language models and token-processing ideas, relationship to masking, preferred description of empathy, and willingness to use computational metaphors. The full conversations are published on the accompanying interview corpus page. These are not substitutes for fieldwork. Their purpose is closer to a wind tunnel: construct sharply different cases, expose a conceptual design to them, and see where it bends.
The method is autoethnographic in a narrow, reflexive sense. The object under observation is not “autistic culture” produced from simulated testimony. It is my own interpretive machinery in contact with a deliberately adversarial set of constructed voices. After each interview I record what I expected, what I tried to make fit, and what distinction the encounter forced me to add or abandon.
4. Six ways empathy refused to be one thing
Mara, a backend engineer deeply familiar with language models, described empathy as building a model of what a situation is costing another person while preserving uncertainty about whether that model is correct. The surprising part was not the model metaphor but the ethic attached to it: “If the stakes are high, I ask.” Her technical vocabulary made explicit inference look less like coldness than a refusal to confuse an internal hypothesis with another person.
Dev, an accessibility engineer with little interest in computational theories of mind, described nearly the reverse problem. Other people’s distress could become physically contagious, but strong resonance reduced the capacity to perform the calm, synchronized behavior that observers expected. “If my output is flat because I am saturated, they infer I am empty.” This was the first moment my framework broke productively. I had separated feeling from inference but had not separated either from expressive bandwidth.
Inez, a compiler engineer, relocated empathy into the relation itself. Explicit questions such as “Are you asking me to listen or solve?” were not evidence of failed understanding; they were repair mechanisms. The difficulty was that one person could experience direct clarification as care while another experienced the need for clarification as proof that care was absent. The interaction contained two sincere attempts at understanding and one culturally privileged standard for what successful understanding should look like.
Noah, an SRE with a long history of masking, destabilized behavior even further. He had learned highly successful social scripts and could produce responses that others judged empathetic while feeling internally disconnected. At other times, strong concern produced atypical or delayed expression and was judged cold. Prediction and performance, he insisted, could support care but could also support masking, manipulation, or simple compliance. Behavioral fluency was therefore not a transparent window onto empathy.
Lin, a machine-learning researcher, pushed the computational analogy toward epistemic humility. Human beings, she argued, learn social priors and then mistake them for direct perception. Her own need to “query more” could be costly in a culture that rewards immediate confidence, yet asking could also protect the other person from confident stereotyping. Ruth, a database engineer largely uninterested in computational accounts of mind, supplied the negative case. She sometimes could not identify another person’s precise emotion but would alter work, bring food, absorb a pager shift, or remove demands. “If empathy requires me to generate the correct emotion inside myself first, then maybe I am not empathetic. But that definition seems morally backward.”
I had expected computational exposure to divide the sample. Instead, it divided the vocabulary. The deepest differences concerned uncertainty, intensity, timing, display, and moral response. Technical concepts helped some interlocutors articulate those differences, but they did not determine them.
5. The latency of care
The strongest synthesis concerns time. Social judgments of empathy are frequently made under tight temporal expectations. A facial expression should be recognized quickly; the appropriate tone should appear without deliberation; reassurance should arrive in the expected sequence; a caring person should supposedly know what to do without asking. Empathy is therefore not merely assessed by content. It is assessed by latency.
This matters because the constructed interviews repeatedly separate fast detection from slower interpretation. Mara knows immediately that something is wrong but delays action while differentiating among possible responses. Dev feels immediately but may lose the ability to display that feeling. Inez can repair uncertainty through explicit questions, but the question itself can be judged too late or too mechanical. Lin argues that slower inference can sometimes be more calibrated precisely because it remains revisable. In each case, the same temporal feature can be read either as impairment or as care depending on which social norm evaluates it.
The anthropological claim is not that autistic empathy is inherently slow. That would simply replace one stereotype with another. The claim is that cultures contain tacit timing conventions for what recognizable empathy looks like, and those conventions can produce false negatives. A delayed, explicit, or differently sequenced response may be interpreted as absence even when substantial affective resonance, inferential work, concern, or practical care is present.
Empathy is socially judged not only by whether understanding or care occurs, but by whether it arrives through the expected channel and within the expected time. Apparent empathic deficit can therefore emerge from a mismatch between internal process and culturally recognized display.
6. Empathy as a stack rather than a score
The interviews suggest a practical decomposition, but not the one I began with. Empathy is better treated here as a stack of partially independent events: another person’s state produces some degree of affective resonance; cues support one or more interpretations; the observer assigns confidence to those interpretations; concern or obligation may arise; an action is selected; and the action becomes more or less legible to the recipient. Each stage can vary without moving in lockstep with the others.
This explains otherwise puzzling combinations. High resonance can coexist with poor outward expression. Accurate inference can coexist with low emotional contagion. Strong moral concern can coexist with uncertainty about the other person’s internal label. Excellent social prediction can coexist with masking rather than care. And a caring action can be illegible if it violates expectations about eye contact, tone, timing, verbal reassurance, or physical proximity.
The point is not to turn empathy into an engineering diagram. It is to stop using one observable surface as a proxy for an invisible whole. Once the stack is visible, “does this person have empathy?” becomes a much less informative question than “what happened between perception, interpretation, concern, response, and recognition in this particular encounter?”
7. What computer-science culture contributes
The technical sample matters, but not because programmers are secretly more computational human beings. Computer-science practice supplies a public vocabulary for processes that other cultures may leave tacit: uncertainty, state, recursion, debugging, interfaces, priors, failure modes, tests, and repair. For some autistic practitioners, this vocabulary may make socially costly experiences easier to externalize. “I failed to empathize” can become “I had three plausible interpretations and no safe way to resolve them,” or “I recognized the state but had insufficient expressive bandwidth,” or “the expected response script conflicted with the response I believed would actually help.”
That vocabulary can liberate, but it can also deceive. Noah’s scripts show that technicalized social reasoning can become camouflage. Ruth’s resistance shows that technical employment does not entail a desire to computationalize intimate life. The anthropological object is therefore not a computational essence of autism. It is a technical subculture that gives some people unusual permission to make the hidden intermediate work of social understanding explicit.
8. The difference between empathy and legibility
The most important distinction in the paper is between empathic process and empathic legibility. The first concerns whatever combination of resonance, inference, concern, and action occurs within and between people. The second concerns whether those processes produce signs that another person recognizes as empathy.
Legibility is culturally organized. A workplace may recognize immediate verbal reassurance but not quiet task removal. A family may recognize physical affection but not information gathering. One neurotype may interpret direct questioning as respectful calibration while another interprets it as evidence that the questioner “does not get it.” When legibility is mistaken for empathy itself, dominant communication norms become invisible and the person who departs from them carries the entire explanatory burden.
This is where the double-empathy problem becomes especially useful. Mutual misunderstanding is not symmetrical in consequence when one communication style has institutional authority. Both people may misread each other, but only one is likely to have the misreading translated into a diagnostic or moral deficit. Anthropology contributes by making that asymmetry visible.
9. What changed in me
I entered the project interested in the variance of deliberate reasoning: how many intermediate representations a person can hold, how recursively they can inspect them, and when they decide not to trust the first answer. I expected empathy to become an application of that idea. It did, but only partially. The interviews made the more interesting object the social judgment surrounding explicit cognition.
I had implicitly treated automatic social understanding as the baseline and deliberate social understanding as compensation. By the end of the exercise, that hierarchy no longer held. Automatic inference can be elegant, efficient, and accurate; it can also be confidently wrong. Deliberate inference can be effortful and slow; it can also create room for uncertainty, consent, and correction. Neither route guarantees care. Neither route guarantees accuracy. What differs is how visible the work becomes and how a culture values that visibility.
The deepest shift was from a psychology of possession to an anthropology of coordination. I began by asking who has empathy and how much. I ended by asking how two people establish enough shared meaning for care to become actionable and recognizable. That is a smaller claim than a theory of autistic empathy and a more useful one.
10. Implications for real fieldwork
A real interview study should therefore avoid asking participants to rank their empathy globally. It should collect concrete episodes and reconstruct their temporal sequence: what was noticed first, what was felt, what interpretations appeared, what uncertainty remained, whether clarification was possible, what action followed, how long each stage took, and how the other person interpreted the response. Interviews should also include episodes of being misunderstood, because the same event can reveal both empathic process and empathic legibility.
The computational-exposure variable should be retained, but as a linguistic and cultural variable rather than a causal hypothesis. The question is whether familiarity with technical models changes what can be said about social cognition, not whether neural-network concepts make someone more or less empathic. The strongest empirical version of this project would also allow autistic participants to revise the analytic categories, identify negative cases, and contest the researcher’s interpretation.
Conclusion
The synthetic interviews do not establish facts about autistic computer scientists. They establish a better problem. Empathy is not a scalar hidden inside an individual and transparently revealed by conventional social behavior. It unfolds across feeling, inference, uncertainty, concern, action, timing, and recognition. Computer-science cultures can make some of these intermediate processes unusually speakable, while autism makes the politics of their legibility difficult to ignore. The resulting thesis is simple: care can be present before it becomes legible, and sometimes the delay between the two is precisely where careful understanding is happening.
References
- Anderson, L. (2006). Analytic autoethnography. Journal of Contemporary Ethnography, 35(4), 373–395. doi:10.1177/0891241605280449
- Bird, G., & Cook, R. (2013). Mixed emotions: The contribution of alexithymia to the emotional symptoms of autism. Translational Psychiatry, 3, e285. doi:10.1038/tp.2013.61
- Cook, R., Brewer, R., Shah, P., & Bird, G. (2013). Alexithymia, not autism, predicts poor recognition of emotional facial expressions. Psychological Science, 24(5), 723–732. doi:10.1177/0956797612463582
- Milton, D. E. M. (2012). On the ontological status of autism: The ‘double empathy problem’. Disability & Society, 27(6), 883–887. doi:10.1080/09687599.2012.710008
- Heasman, B., & Gillespie, A. (2019). Neurodivergent intersubjectivity: Distinctive features of how autistic people create shared understanding. Autism, 23(4), 910–921. doi:10.1177/1362361318785172
- Crompton, C. J., Ropar, D., Evans-Williams, C. V. M., Flynn, E. G., & Fletcher-Watson, S. (2020a). Autistic peer-to-peer information transfer is highly effective. Autism, 24(7), 1704–1712. doi:10.1177/1362361320919286
- Crompton, C. J., Sharp, M., Axbey, H., Fletcher-Watson, S., Flynn, E. G., & Ropar, D. (2020b). Neurotype-matching, but not being autistic, influences self and observer ratings of interpersonal rapport. Frontiers in Psychology, 11, 586171. doi:10.3389/fpsyg.2020.586171
- Kimber, L., Verrier, D., & Connolly, S. (2024). Autistic people’s experience of empathy and the autistic empathy deficit narrative. Autism in Adulthood, 6(3), 321–330. doi:10.1089/aut.2023.0001
- Jones, D. R., Botha, M., Ackerman, R. A., King, K., & Sasson, N. J. (2024). Non-autistic observers both detect and demonstrate the double empathy problem when evaluating interactions between autistic and non-autistic adults. Autism. doi:10.1177/13623613231219743
- Márquez, G., Pacheco, M., Astudillo, H., Taramasco, C., & Calvo, E. (2024). Inclusion of individuals with autism spectrum disorder in Software Engineering. Information and Software Technology, 170, 107434. doi:10.1016/j.infsof.2024.107434