Synthetic corpus · qualitative method probe
Six Interviews About Empathy
Distinct constructed conversations with autistic computer-science practitioners, designed to pressure-test the paper’s assumptions rather than impersonate human-subject data.
The corpus was constructed to resist a single story. The six interlocutors differ in technical role, prior exposure to neural-network and token-processing ideas, preferred language for empathy, relationship to explicit social reasoning, and attitude toward computational metaphors.
Interview 1 · “Mara”
Calibration, not mind-reading
Author: When you say you empathize with someone, what do you mean?
Mara: Usually that I have built a model of what the situation is costing them, and that I care about that cost. I do not mean I have reproduced their feeling inside myself. Sometimes I do. But those are separate things.
Author: Does “model” make it sound colder than it feels?
Mara: Only if you think unconscious guessing is warmer than conscious guessing. People are modeling each other constantly. I am just less comfortable pretending my model is the person. If the stakes are high, I ask.
Author: So explicit reasoning helps?
Mara: It helps me preserve uncertainty. The dangerous moment is when I get one interpretation and it feels finished. In code review, one failing test does not tell you the root cause. Socially, people often act as if one facial expression does.
Author: Do you experience empathy as slow?
Mara: Recognition can be fast. Confidence is slow. I might instantly know something is wrong and need ten minutes to know whether the right response is a question, practical help, silence, or leaving them alone. Other people sometimes read that latency as not caring. It is almost the opposite. I am waiting because I care about being wrong.
I had expected technical vocabulary to produce a mechanistic account. Instead, Mara used model language to argue against epistemic overconfidence. Her computationalism did not replace care; it disciplined inference.
Interview 2 · “Dev”
Too much feeling, too little bandwidth
Author: What is empathy like for you?
Dev: Physical, annoyingly. If somebody is embarrassed, I feel heat in my face. If somebody is panicking, I can become useless because my own body starts doing the same thing. I used to think that meant I was good at empathy. Then I realized feeling a lot does not tell me what the other person needs.
Author: Do you reason through it afterward?
Dev: Afterward, yes. During it, reasoning can disappear because I am overloaded. The socially expected thing is to look calm, say the right sentence, keep eye contact, maybe touch someone. Those are exactly the things that can become harder when I am feeling the most.
Author: Would a processing metaphor help describe that?
Dev: Maybe bandwidth. But I would not want you to turn me into a router. The important thing is that people see the output and infer the inside. If my output is flat because I am saturated, they infer I am empty.
Author: Is that a misunderstanding of empathy?
Dev: It is a misunderstanding of expression. People say empathy when they often mean “performed reassurance in the dialect I recognize.”
Dev broke my first decomposition. I had separated feeling from inference, but not feeling from the capacity to display feeling. The interview forced a third variable into view: expressive bandwidth.
Interview 3 · “Inez”
The interface is part of the phenomenon
Author: Have you ever felt that you understood another autistic person more easily?
Inez: Often, but not magically. It is more that some defaults line up. Someone can say, “I cannot talk now; send me the exact question,” and I do not hear rejection. I hear a useful interface.
Author: Interface sounds computational.
Inez: It is, but the point is relational. An interface only exists between things. If two people keep failing to communicate, asking which one has the broken empathy module is already the wrong level of analysis.
Author: What happens in mixed interactions?
Inez: I can spend a lot of effort decoding implications that the other person thinks are obvious. They spend effort decoding my literalness or lack of facial signaling. Both of us may be trying. But only one style gets called normal, so only one person gets diagnosed as failing.
Author: Does explicitness improve things?
Inez: Usually. “Are you asking me to listen or solve?” is an incredible sentence. But some people experience explicit questions as unempathetic because they expect you to know. That is the trap: asking can be evidence that I care enough not to guess, but it can be read as evidence that I do not understand.
Inez shifted the unit of analysis from person to relation. The most consequential variable was not how much empathy either person possessed, but whether their signals were mutually interpretable and whether explicit repair was socially permitted.
Interview 4 · “Noah”
Prediction can become performance
Author: You described learning social responses almost like tests. What did you mean?
Noah: I learned rules. If someone tells you about a death, do not optimize the conversation, do not compare it to your own story, acknowledge the loss, leave space. Those rules helped because my first impulse might be practical. But eventually I got too good at passing the tests.
Author: Too good?
Noah: I could produce the expected response while being completely disconnected from myself. People would call me empathetic because the output looked right. On days when I actually felt something strongly but could not perform, they would call me cold. That made me suspicious of the whole category.
Author: So social prediction is not empathy?
Noah: It can support empathy. It can also support sales, manipulation, masking, threat detection. Predicting someone is not the same as caring about them. The moral direction matters.
Author: What changed after diagnosis?
Noah: I started replacing performance with negotiated clarity. I say, “I care, and I am not sure what response would help.” That sentence feels more honest than a perfect script. Weirdly, some people trust it more.
Noah separated social competence from empathy more sharply than I had. Accurate prediction can produce caring action, but it can also produce camouflage. Behavioral legibility is therefore an unreliable proxy for inner concern.
Interview 5 · “Lin”
The ethical value of uncertainty
Author: Does working with models change how you think about empathy?
Lin: It mostly makes me distrust confident interpretation. Models can be very good on average and very wrong about an individual. Humans do the same thing with social categories. They learn a prior and then call it intuition.
Author: Is empathy a prediction problem?
Lin: Partly. But if you define it that way, the best empath becomes the best classifier, which is obviously wrong. Empathy includes what you do with uncertainty and what obligations you accept toward the person you are modeling.
Author: What does autism change in that picture?
Lin: For me, some social priors are weak or expensive. I have to query more. But that can become a strength if I am allowed to query. I ask the person instead of treating the stereotype as ground truth. The social world often rewards low-latency confidence, not calibrated accuracy.
Author: Would you call that System-2 empathy?
Lin: I would avoid making a hierarchy. Deliberate empathy is not superior to intuitive empathy. It is just visible to consciousness in a different way. And deliberate processes can become fast with familiarity. The more interesting question is who gets penalized for needing the explicit route.
Lin returned the project to its starting question about reflective cognition, but inverted its value. Slow processing was not merely a cognitive limitation to be overcome; in some contexts, latency created room for calibration, consent, and correction.
Interview 6 · “Ruth”
Care without simulation
Author: Do you feel other people’s emotions?
Ruth: Sometimes. More often I notice consequences. Someone is missing deadlines, skipping lunch, making mistakes they do not normally make. I may not know whether the internal label is grief, fear, shame, or exhaustion. I still know something has changed.
Author: Is that empathy?
Ruth: I do not care what category it goes in. I will take the pager shift, bring food, remove a meeting, or say, “You do not have to explain this to me.” If empathy requires me to generate the correct emotion inside myself first, then maybe I am not empathetic. But that definition seems morally backward.
Author: Does your technical background shape that style?
Ruth: Probably less than you want it to. I think you are in love with your metaphor. I have been programming for thirty years and I do not experience people as programs. I experience them as people who are often terrible at stating their requirements.
Author: That sounds like a technical metaphor.
Ruth: Fine. You got me. But do not make the metaphor the finding.
Ruth supplied the negative case the project needed. Technical work did not imply enthusiasm for computational self-description. Her account relocated empathy from accurate inner simulation to accountable action, and her resistance exposed my desire to make the corpus cohere too neatly.
What the constructed corpus was designed to vary
The interviews deliberately separate dimensions that deficit narratives often collapse: affective resonance, inference about another person’s state, confidence in that inference, moral concern, behavioral response, expressive legibility, processing latency, willingness to ask directly, and prior exposure to computational models. No single synthetic participant is intended to stand for an autistic “type.” The useful object is the contrast among them.