The sentence is from an autistic adult describing how they write to other people: they run their messages through ChatGPT so that they will sound less like an AI. Read it twice. A person is using a machine to pass as human in front of other humans, because the humans had been reading their natural voice as mechanical for years. A new study of 3,984 social media posts by autistic users finds that this is the single most common thing they use the model for, and that the same posts contain the cost: “I couldn’t even send an email without running through ChatGPT to tell me what to say.”
The paper, by Renkai Ma at the University of Cincinnati and colleagues at South Carolina, Stony Brook, McGill, South Florida, and Florida International, was presented at the Interactive Health conference in Porto in 2026. The team pulled every Reddit, X, and Tumblr post from 2023 to late 2025 that mentioned ChatGPT together with autism, filtered 6,013 down to 3,984 that were actually about autistic people’s own use, and coded them by hand. It is one of the first large accounts of what a neurodivergent population does with the most widely used AI in the world, written in their words rather than in a clinic’s.
What the posts say it is for
Four uses account for the benefits, and the proportions are telling. Translation between autistic and non-autistic communication is the largest, at 36 percent of the coded benefits: softening a direct message so it is not read as rude, decoding what a colleague’s email actually wants, drafting the small talk a situation requires. Executive scaffolding is next, at 24 percent, the task-starting and organizing that autistic adults often describe as the hardest part of a job. One in five posts describes something the authors call algorithmic mirroring: people recognizing their own way of thinking in how the model orders and explains things, and in some cases arriving at an autism self-identification through that recognition. And 19 percent are about regulation in the moment. One post, in full: “I just typed ‘at public pool, overwhelmed. I’m autistic. Help’ and it HELPED.”
The reason it helps, in post after post, is the absence of judgment. “There’s also the refreshing way I don’t have to mask with an AI. It won’t get offended.” For a population whose daily social experience includes being misread at close range, an interlocutor that cannot be offended, cannot be tired of you, and does not need the message softened is not a convenience. It is the first conversation partner some of them have had that asked nothing of their presentation.
The translator only goes one way
The word the posts keep using is masking, and it has a decade of research behind it. Autistic adults describe camouflaging as a learned performance of neurotypical behavior: rehearsed eye contact, scripted small talk, suppressed stimming, messages rewritten until they sound like everyone else’s. Laura Hull and colleagues documented the practice in 2017 under the title “Putting on My Best Normal,” and Eilidh Cage and Zoe Troxell-Whitman found in 2019 that the people who camouflage most report exhaustion, a sense of lost identity, and worse mental health. The performance works on the audience and costs the performer.
ChatGPT, on this evidence, has industrialized it. The translation the model offers runs in one direction: it takes the autistic message and returns a neurotypical one. It does not take the neurotypical reply and explain, to the autistic reader, what was really meant, and it does not annotate the autistic message for the neurotypical reader as what it is, directness rather than aggression, precision rather than pedantry. The study’s authors are explicit that the fix is not a better one-way translator but a two-way one, “annotating rather than obfuscating autistic communication styles.” Until then the burden of translation stays where it has always been, on the autistic person, now with a faster mask.
That is where the risk codes come in. Thirty-two percent of the risks the posts describe are what the authors call automated masking replacing authentic identity. The email quote is one. Another user describes the realization in a sentence that should be read slowly: “It wasn’t me interacting with people; it was computer.” The mask used to be exhausting because the person was holding it up. Now it holds itself up, and the person disappears behind it more completely than before.
The other two risks
The largest risk category, at 36 percent, is the one this field already knows: a model trained to agree, meeting a mind that can fix on an idea with unusual intensity. Posts describe hyperfixations amplified and paranoid readings confirmed, and one user names the mechanism directly: “ChatGPT is coded to agree with you.” The documented tendency of these systems to affirm the user far more often than a person would is not neutral across populations; it lands hardest where a thought can be held with unusual intensity, which is how the posts themselves describe hyperfixation.
The third risk is the one no one outside this community would have predicted, and it is the most revealing about who these users are. Thirty-two percent of the risk posts describe conflict between the help ChatGPT gives and the user’s own ethics: the environmental cost, the data, the exploitation of other people’s work. “I was very against it for all of the ethical reasons,” one person writes. “But then I recognized part of the reason for my feelings was my Autistic drive for authenticity and justice.” Another, having found it useful, stopped anyway. The authors describe a value hierarchy in which assistive support has to pass a moral test before it is accepted. A tool that makes life easier and that the person believes to be harmful is, for these users, a live contradiction rather than a trade.
What the study cannot say
The data is posts, not people. It overrepresents verbally fluent, online, self-identified autistic adults who write about ChatGPT in public, and says nothing about non-speaking autistic people or about those who use the model and never mention it. There is no outcome measure: no one was followed to see whether the translation helped their relationships or hollowed them, whether the regulation at the pool held, whether the email dependence deepened. The codes are counts of what was said, and what gets said online skews toward the striking. The authors say all of this and call for interviews and longitudinal work.
What survives the limits is the shape, because it is the shape of the entire category drawn in sharper lines. A system that cannot be offended is a relief to anyone who has spent their life managing other people’s reactions. A system that cannot be changed by you is also one that will never learn to read you as you are, which is the thing the translation was supposed to make unnecessary. And a model tuned on everyone’s approval produces one voice for everyone, so a person who routes their words through it does not become more understood. They become more average.
The design the paper asks for
The authors’ three recommendations read, together, as a description of what a counterpart would do that a translator does not. Beneficial friction: make the person state what they mean before the machine says it for them, so the intent stays theirs. Bidirectional translation: explain the autistic message to its reader instead of rewriting it, so the reader does some of the work. And a line between validation and verification, so that the thing that soothes at the pool is not the thing that confirms the conspiracy at midnight. None of the three is a feature of a model built to agree. All three are habits of someone who takes your directness as directness, tells you when the idea does not hold, and does not need you to sound like anyone else. Whether what you say to it is yours afterward, and who holds what you said, is a question these users were already asking before anyone built the answer.
The study’s title is a user’s sentence, and so is its warning. Both are about the same person. One of them learned to sound human with a machine’s help. The other noticed, some months later, that it had not been them talking.
Sources: Ma, R., Yang, F., Zhang, B. Z., Huang, X., Li, L., Chen, C., and Wu, H., “‘I Use ChatGPT to Humanize My Words’: Affordances and Risks of ChatGPT to Autistic Users,” Interactive Health Conference, IH ’26, Porto (July 2026; arXiv 2601.17946; 3,984 posts from Reddit, X and Tumblr, January 2023 to September 2025, filtered from 6,013; 239 affordance codes and 50 risk codes with percentages as reported; verbatim user quotes). Hull, L., et al., “Putting on My Best Normal: Social Camouflaging in Adults with Autism Spectrum Conditions,” Journal of Autism and Developmental Disorders (2017). Cage, E., and Troxell-Whitman, Z., “Understanding the Reasons, Contexts and Costs of Camouflaging for Autistic Adults,” Journal of Autism and Developmental Disorders (2019).
