I asked it to translate my boss. It taught me to pretend better.

Artificial intelligence can ease a social burden that should never have rested solely on autistic people. The question is whether it also helps share that burden.

Imagine the scene. Eleven at night. An autistic researcher with low vision listens through her screen reader to an email that arrived at six that afternoon. She has returned to those two sentences a dozen times:

“We need to talk about the report. Come see me first thing tomorrow.”

There is no specific subject, no indication of urgency, no context. It could be a routine adjustment, praise or a reprimand. She does not know. She keeps trying to figure it out and cannot sleep.

She opens a chatbot, pastes the message and asks what it means. The answer organizes three interpretations and suggests it is probably something practical. She exhales. The machine seems to have understood her boss.

Then she asks for help replying. The suggestion comes:

“Of course! I’d be happy to chat. See you tomorrow!”

It has enthusiasm and an “I’d be happy” that she does not feel. She sends it anyway because it sounds the way people expect it to sound.

On that imagined night, she sleeps more peacefully. Her boss’s intention remains unknown. And another question remains: who did the work of making this conversation possible?

1. The tax nobody collects out loud

Every conversation seems to come with a manual nobody hands out. How much enthusiasm to add, when a short answer sounds offensive, how to disagree without seeming aggressive. The content can be correct and the person can still leave the interaction owing an explanation for their way of speaking.

Erving Goffman described the management of the impressions we produce in everyday encounters. In the 1959 edition of The Presentation of Self in Everyday Life, the theater metaphor helps explain this work: we present a version of ourselves and try to sustain the scene in front of others. Goffman, 1959 [1].

Harold Garfinkel showed how social life depends on tacit expectations. In an exercise reported in Studies in Ethnomethodology, students began behaving at home as guests. The strangeness emerged precisely where there had previously seemed to be only a family’s “natural” behavior. The rule became visible when someone stepped outside it. Garfinkel, 1967, chapter 2 [2].

For many autistic people, deducing and carrying out these expectations requires conscious effort. Laura Hull and colleagues’ research with 92 autistic adults revealed camouflaging strategies, motivations linked to acceptance and accounts of exhaustion and threats to a sense of identity. The paper’s title captures the discomfort: Putting on My Best Normal. Hull and colleagues, 2017 [3].

I am not saying every adaptation of language is a mask. Choosing words is part of communication. The distinction appears when choice becomes a constant requirement to hide one’s own characteristics in order to be treated with respect.

The cost deserves to be taken seriously. Dora Raymaker and colleagues described autistic burnout in participatory research: prolonged exhaustion, loss of skills and reduced tolerance to stimuli in contexts where demands exceed available resources and support is insufficient. Raymaker and colleagues, 2020 [4].

A 2025 systematic review brought together 48 studies involving approximately four thousand autistic people. Chronic camouflaging, stigma and daily demands appear among factors related to burnout. The combined sample, however, primarily represents white adults, women and people diagnosed later, with at least average intellectual or verbal abilities. The experience of the entire spectrum does not automatically fit within that sample. Ali and colleagues, 2025 [5].

Saving effort matters. Claiming that AI prevents burnout would require a different kind of evidence.

2. Difficulty understanding does not belong to only one side

An influential line of research treated autistic social difficulties as a lack of “theory of mind”: the ability to attribute mental states to other people. Morton Ann Gernsbacher and Melanie Yergeau’s critical review questions the empirical basis for this lack as a universal and autism-specific characteristic. This does not erase individual difficulties; it changes the scope of the explanation. Gernsbacher and Yergeau, 2019 [6].

In 2012, Damian Milton proposed the double empathy problem. Instead of locating every mismatch in the autistic person, the proposal examines reciprocity: people with different experiences and ways of perceiving may struggle to understand one another. A non-autistic person’s interpretation can fail too. Milton, 2012 [7].

In 2020, Catherine Crompton and colleagues studied stories passed along chains of participants, like a game of telephone. Autistic chains transmitted information comparably to non-autistic chains; mixed chains experienced greater loss in that experiment. Crompton and colleagues, 2020 [8].

There is an important update. In 2025, an expanded replication with 311 participants across three sites found no significant difference in transmission between the three chain types, including mixed chains. The selective disadvantage in mixed chains did not replicate. The conclusion needs to preserve both findings: effective communication occurred between autistic people, and mismatch between neurotypes is not inevitable in every task. Crompton and colleagues, 2025 [9].

A story-transmission task also does not reproduce all the pressure of interpreting a boss. There is hierarchy, professional consequences, relationship history and rules that may change without warning.

In anthropology, Elinor Ochs and Olga Solomon discussed “autistic sociality” through research with children: social coordination depends on partners, activities and environments, and can take diverse forms. This perspective helps us ask about the conditions of interaction. Ochs and Solomon, 2010 [10].

Return to the email. Its writer did not explain the subject. Its recipient spent the night trying to guess it. My question concerns that distribution: why is clarity from the writer often optional, while the recipient’s ability to interpret becomes an obligation?

3. The channel changes the conversation

Text can provide time to think, review and choose when to respond. For some people, this reduces the pressure of managing voice, face, gestures and content simultaneously. For others, text creates barriers. No single interface represents every autistic person’s preferences.

Kristen Gillespie-Lynch and colleagues compared perceptions of online communication among 291 autistic and 311 non-autistic participants. The first group reported benefits such as greater control over interaction and greater ease expressing aspects of themselves. These are participants’ perceptions from an online survey, not a rule of preference for the entire spectrum. Gillespie-Lynch and colleagues, 2014 [11].

Philippa Howard and Felicity Sedgewick investigated 245 autistic adults’ preferences across different situations. The title Anything but the phone! hints at the result: the phone was particularly unwelcome when dealing with unfamiliar people, while writing was valued. Preferences also depended on context and the communication partner. Howard and Sedgewick, 2021 [12].

There is also Noah Sasson and colleagues’ 2017 work. Non-autistic observers formed less favorable impressions of autistic people from brief presentations. In the condition showing only a transcript of speech, the difference between groups was not statistically significant. Sasson and colleagues, 2017 [13].

That result points to the importance of presentation cues in those evaluations. It does not show that all prejudice disappears in text, or that all content produced by two groups has identical quality. It shows that social judgment can change when the channel changes.

For me, this discussion has a concrete layer: I am autistic and have low vision. I use a screen reader and a voice-to-voice interface every day. Accessible text can be heard; a screen full of text in unlabeled controls can be impossible to use. “Preferring writing” and “being required to see an interface” are different things.

AI is part of my accessibility. That is the position from which I want to examine it.

4. The relief exists — and we need to understand what was measured

Jeffrey Hancock, Mor Naaman and Karen Levy define AI-mediated communication as communication in which a system modifies, augments or generates messages on someone’s behalf. That definition shifts the focus: beyond talking to a machine, we start talking to other people through it. Hancock, Naaman and Levy, 2020 [14].

It can organize thoughts, offer versions of an email and help explore interpretations. That does not give the model access to the writer’s intention. In the opening email, the answer’s confidence can reassure without resolving the ambiguity.

In JiWoong Jang and colleagues’ study, presented at CHI 2024, eleven autistic participants discussed workplace situations with a GPT-4-based chatbot and with a human partner presented as a chatbot. Nine preferred the interactions with the model. This signals perceived value in a specific situation with a small sample. Jang and colleagues, 2024 [15].

A professional counselor and career advisor who evaluated the responses raised concerns about unsupported assumptions and recommendations centered on neurotypical norms. That difference in perspective is part of the result: perceived usefulness and advice quality need to be examined together, without erasing the preference of the people using the tool.

Buse Carik and colleagues analyzed discussions from 61 neurodivergent communities on Reddit in work associated with GROUP 2025. They found uses in communication, organization, learning and emotional support, alongside criticism of excessively neurotypical responses and concerns about dependence. The sample includes different forms of neurodivergence; it is not an exclusively autistic clinical sample. Carik and colleagues, 2025; 2024 manuscript [16].

In 2026, Renkai Ma and colleagues brought the two sides even closer: support and automated camouflaging. Their analysis of 3,984 social media posts describes benefits and tensions in accounts from people identifying as autistic. The manuscript indicates acceptance as a poster/extended abstract at Interactive Health 2026. This is preliminary research on posts, which do not equal 3,984 clinically followed participants. Ma and colleagues, 2026 [17].

These studies help explain why someone seeks out the tool and what they find there. To evaluate lasting effects, we also need to observe what changes in the person’s life after the conversation ends.

5. Outsourcing the mask does not necessarily mean taking it off

Remember “I’d be happy to chat”? The answer may have solved the immediate task while hiding the sender’s state. If a tool can only make someone acceptable by replacing their voice with a required display of enthusiasm, it is worth asking which problem was solved.

Arlie Hochschild analyzed emotional labor in The Managed Heart, from 1983: managing feelings and their expression as part of paid work. In 2026, Young Ji Kim and Margot Plunkett, through interviews with 31 autistic workers, developed the concept of assimilation labor: deliberate, sustained and often invisible efforts to adapt to normative expectations. Hochschild, 2012 edition of the 1983 book [18]; Kim and Plunkett, 2026 [19].

The second study does not test AI. I am proposing the connection: some of this effort may start being done with a model’s help, while the organizational demand remains unchanged.

That can save energy and expand autonomy. One person may choose a warmer version and still recognize themselves in it. Another may feel they need to disappear inside the accepted version. The criterion cannot simply be fluent text: choice, cost and recognition of one’s own voice need consideration.

We also need to investigate the tool’s criteria. In AUTALIC, presented at ACL 2025, Naba Rizvi and colleagues assembled 2,400 sentences about autism with context and annotations to evaluate ableist language. The evaluated models struggled to identify that ableism reliably. A system helping mediate communication also needs evaluation of the prejudices it lets through. Rizvi and colleagues, 2025 [20].

In 2025, ASAN published a position against using generative AI to produce plain language. This is a self-advocacy position, distinct from an experimental trial, and belongs in the discussion of who defines the quality of accessibility. ASAN, 2025 [21].

To me, a sentence being easier to accept socially does not, by itself, make it better for the person who needs to use it.

6. The questions that remain open

In experiments with automatic reply suggestions, Jess Hohenstein and colleagues observed faster communication and more positive language. They also found less favorable evaluations when participants perceived their partner as making greater use of algorithmic replies. These studies were not an autism-specific evaluation or a test of text detectors. Hohenstein and colleagues, 2023 [22].

From there, I raise a hypothesis: could people whose writing is perceived as formal, direct or unconventional be falsely accused of using AI? And could someone using an accessibility aid face a second penalty for relying on it?

There is no single “autistic writing style,” and the cited studies do not demonstrate this chain of effects. It is a question I would like to investigate, distinguishing human suspicion, detector performance and real consequences for writers.

Concern about replacing human relationships also appears in accounts about models. It deserves follow-up but does not permit diagnosing dependence from frequent use. A tool may complement a support network, open a previously impossible conversation or occupy a space left by inaccessible alternatives. We need to understand the conditions and differences between those situations.

The literature offers concepts and research on autism, interaction, stigma and technology. The contribution I want to build brings them closer to concrete situations of AI mediation and my experience of accessibility. I want to understand who makes the effort, who defines an answer as appropriate and who can reject that definition.

In Brazil, Francisco Ortega discussed relationships between identity, brain-centered explanations and the neurodiversity movement in 2008. It is a critical analysis of those relationships, not a study of today’s chatbots. Ortega, 2008 [23].

Monique Botha and David Frost, in work published in the 2020 volume of Society and Mental Health, found an association between minority stressors and poorer mental health in a sample of 111 autistic people. Their contribution considers discrimination, stigma and concealment when explaining distress, alongside individual characteristics. Botha and Frost, 2020 [24].

These are foundations for better questions. They do not replace studies of a specific tool, population and outcome. Nor does a result from a 2024 chatbot automatically describe every 2026 model: systems change, interfaces change and people remain different from one another.

7. What I ask of a tool

I want the effort of communication to be shareable. In the opening scene, the assistant could help the researcher ask for context. It could also help her boss notice what was left unexplained.

Instead of guessing on the boss’s behalf, it could offer her:

“The email does not explain the reason for the conversation. I can organize possibilities, but cannot determine the writer’s intention.”

One possible reply, without manufactured enthusiasm, would be:

“OK, see you tomorrow. Could you tell me which part of the report we’ll discuss? That way I can prepare.”

On the other side, the tool could suggest to the sender: “Would you like to state the subject and whether any preparation is needed?” The boss would fill in what is true. The model would not invent an assurance that everything is fine.

These examples are design proposals. Accessibility criteria also need to appear in how the tool works:

• Explain before rewriting. Separate what the message says from possible interpretations and missing information.

• Preserve voice and choice. Offer versions, show changes and allow the original text to be retained. Kindness need not become compulsory enthusiasm.

• Make uncertainty usable. Help formulate a clarification question when intention is unknown, instead of simply reassuring with confidence.

• Share the adaptation. Support both recipient and writer, without assuming the majority standard is always right.

• Be accessible from end to end. For me, that includes screen readers, understandable navigation and voice, with freedom to choose the channel and pace.

• Provide control over data. Make retention, access and secondary uses clear; allow information to be removed before sending. In my proposal, intimate messages and other people’s data do not become training material or consumer profiles by default.

I would evaluate these choices with the people who use them: could they say what they wanted? Did they understand what changed? Did it take less effort? Did they feel they remained present in their own message? A response rated highly by its recipient does not answer all those questions.

AI can be important support without resolving, on its own, the norm that makes that support necessary. I want to use the tool, choose a reformulation when it helps me and reject one when it erases me.

That night, the researcher needed rest and information. Tomorrow she may need a boss who explains the subject. I want a world where she can also answer “OK, see you tomorrow” without having to prove affection with an exclamation mark.

For readers who want to go deeper

Interaction, sociality and communication

• Goffman — The Presentation of Self in Everyday Life, 1959 edition [1]. Self-presentation and impression management; the date refers to the indicated English edition.

• Garfinkel — Studies in Ethnomethodology, 1967 [2]. Background expectations in everyday life, especially chapter 2.

• Milton, 2012 — the double empathy problem [7]. A conceptual proposal on reciprocity in understanding.

• Ochs and Solomon, 2010 — Autistic Sociality [10]. An anthropological perspective on social coordination and context.

• Gernsbacher and Yergeau, 2019 — Empirical Failures of the Claim That Autistic People Lack a Theory of Mind [6]. A critical review of the claim of a universal and specific lack.

• Crompton and colleagues, 2020 — information transfer between peers [8], and expanded 2025 replication [9]. Read both findings together.

• Gillespie-Lynch and colleagues, 2014 — computer-mediated communication [11].

• Howard and Sedgewick, 2021 — Anything but the phone! [12].

• Sasson and colleagues, 2017 — first impressions across modalities [13].

Camouflaging, work and burnout

• Hull and colleagues, 2017 — Putting on My Best Normal [3].

• Raymaker and colleagues, 2020 — defining autistic burnout through participatory research [4].

• Ali and colleagues, 2025 — systematic review of burnout [5].

• Hochschild — The Managed Heart, 1983 book, 2012 edition [18]. Emotional labor.

• Kim and Plunkett, 2026 — assimilation labor [19].

• Botha and Frost, 2020 volume — minority stress and mental health [25].

• Ortega, 2008 — O sujeito cerebral e o movimento da neurodiversidade [23].

AI mediation, experiences and critiques

• Hancock, Naaman and Levy, 2020 — AI-Mediated Communication [14].

• Hohenstein and colleagues, 2023 — language and social relationships [22].

• Jang and colleagues, CHI 2024 — communication assistance for autistic workers [15]. Authors’ manuscript, revised in 2025 [26].

• Carik and colleagues, GROUP 2025 — experiences in neurodivergent communities [16]. Manuscript made available in 2024.

• Ma and colleagues, 2026 — I Use ChatGPT to Humanize My Words [17]. Preliminary work, poster/extended abstract accepted at Interactive Health 2026.

• Rizvi and colleagues, ACL 2025 — AUTALIC [20]. Dataset and evaluation of ableist language detection in context.

• ASAN, 2025 — position on generative AI and plain language [21]. An institutional self-advocacy position.

Talk to me

Have you ever asked for help understanding a message and received a reply that did not sound like you? Or found in AI a way to say, at last, exactly what you wanted?

If you want to share that experience privately, write to [email protected]. It can be a short account, including an experience different from the one discussed here. There is no need to send screenshots, names or other people’s data.