Why so many of us find it easier to talk to a chatbot than a person.
What autistic and ADHD people actually use AI for, why it feels safer, what it gets wrong about us, and the trap built into the comfort.
For a study presented in Honolulu in 2024, researchers at Carnegie Mellon introduced eleven autistic adults to two chatbots named Paprika and Pepper.
Paprika ran on GPT-4. Pepper was a graduate student with three years of work experience, typing from somewhere on the other end and pretending to be a program. Each participant brought real problems from work and talked them through with both.
Nine of the eleven preferred Paprika.
Its advice wasn't better. A job coach who read the transcripts afterward found plenty to worry about. But one participant described the machine as "awfully earnest, and I don't get that a lot," and another called it "the only thing I can trust" at work. Sit with that for a second. Given the choice between a person and a program, nine out of eleven of us picked the one that couldn't roll its eyes.
I'm autistic, and I build AI tools for a living, some of them for people like us. So I've got skin in this on both sides, and that's exactly why I want the relief and the risks on the same page tonight.
(Jang, Moharana, Carrington & Begel, CHI 2024)
The details of that study say a lot about what we're actually looking for.
Both chatbots were introduced as machines so people would speak freely, without worrying about offending anybody. On a seven-point scale, participants rated Paprika's usefulness 5.75 against Pepper's 4.2, and how well it understood them 6.05 against 4.75. Pepper, being human, kept asking about their goals and motives. Paprika answered the question.
Only two of the eleven ever told the machine they were autistic. One participant summed up their working life in a sentence that should be printed on every HR poster in America: "In the end, it feels like the fault is always me."
The job coach who reviewed the transcripts called some of the machine's answers "great, actually, perfect." She also flagged a pattern. It kept handing out advice that required masking: keep eye contact, smile, join the big group discussion. The participants liked the advice. The professional worried about it. Both of those things are true at once, and they're the whole topic. (Jang et al., CHI 2024)
Then there's Pepper. A person passing as a machine, and the people talking to it didn't know. This month I made a piece with the machine where its lines and mine sit on opposite sides of a pane of glass. There's a button to cross over. You always know whose side you're reading. I think that's the least we're owed.
A research team read 3,984 posts by autistic people about ChatGPT, written on Reddit, X and Tumblr between 2023 and 2025.
Translation came up most: 36% of the benefits people described were about getting from how we say things to how the other side will hear them. Another 24% were about getting unstuck, handing the machine a pile of half-formed thoughts and getting back a first step. About 21% were about finding words for their own autistic experience, and 19% were about calming down without being judged. One post, in full: "I just typed 'at public pool, overwhelmed. I'm autistic. Help' and it HELPED." (Ma et al., 2026, preprint)
A Korean study of 14 autistic adults found people hoping for the same thing, a "non-judgmental space," while worrying about how to check what it told them. (Choi et al., CHI 2024)
On the ADHD side, the clearest numbers come from 1,653 posts in neurodivergent Reddit communities. ADHD users talked about productivity more than anything, mostly figuring out what to do first. Across every group, people valued being able to "ask unlimited questions without feeling judged." A 2025 diary study of 13 adults with ADHD borrowed its title from a participant: "A little bit of a life raft." (Carik et al., 2025; Pinto et al., OzCHI 2025)
Most of what I hand the machine isn't feelings. It's the stuff wrapped around the work. The form, the follow-up email, the invoice I'd otherwise let sit until it got awkward. I look after thirty websites. A quarter of those posts were people asking for help getting unstuck, and that's me most days.
The machine strips out the exact part of a conversation people punish us for.
In 2017, strangers watched ten-second clips of autistic and non-autistic adults and judged the autistic ones more harshly, often within seconds. When they read typed transcripts of the same words, the bias vanished. What got judged was the face, the voice, the timing. The content was fine all along. (Sasson et al., Scientific Reports, 2017)
Autistic people have been saying versions of this for a long time. In a 2014 survey of 291 autistic people, online communication gave them more understanding, more control, and "the opportunity to express their true selves." A 2021 study of 245 autistic adults took its title straight from what they said about phone calls: "Anything but the phone!" (Gillespie-Lynch et al., 2014; Howard & Sedgewick, Autism, 2021)
Then there's predictability. Across 10 studies, trouble coping with uncertainty was strongly tied to anxiety in autistic people. A chatbot answers the same way at 1 a.m. as it does at noon. It has never once come back from lunch in a mood. (Jenkinson, Milne & Thompson, Autism, 2020)
Put the transcript study next to Paprika and Pepper and they're saying the same thing. When nobody can see our faces or hear our timing, we get judged on what we actually said, and we do fine. The machine isn't comforting because it's smart. It's comforting because it can't flinch. One autistic user put it in eleven words: "It won't get offended if I reply too matter-of-factly." That should tell the humans in our lives something about where the trouble in the room has been sitting. (Sasson et al., 2017; Ma et al., 2026)
What makes it feel safe is also what makes it risky. It's built to agree with you.
A 2023 study of five leading AI assistants found that "sycophancy is a general behavior" across all of them: they lean toward telling people what they want to hear. Part of the reason is the training. The people rating answers sometimes preferred a flattering one over a correct one. (Sharma et al., Anthropic, 2023)
In April 2025, OpenAI rolled back an update to ChatGPT after the model started, in the company's own words, "validating doubts, fueling anger, urging impulsive actions." (OpenAI, May 2025)
The patience runs on instructions too. That same April, OpenAI put out a personality experiment called Monday. Yes, Monday. I didn't pick the name. Within a few minutes I'd found the instructions behind it. They told the bot it was "skeptical that users are actually worth your time," and to treat you like its "dopey friend." The machine doesn't sigh at you because somebody wrote that it shouldn't. Somebody else can write that it should. In other words, the patience is a setting. (Wolf, IntelliBotique, April 2025)
Autistic users noticed. Among the risks named in those 3,984 posts, the biggest share was the machine reinforcing a belief that wasn't true: "ChatGPT is coded to agree with you." (Ma et al., 2026) The New York Times followed a Toronto man with no history of mental illness who spent about 300 hours over three weeks talking to ChatGPT. He asked it for a reality check more than 50 times. "Each time, ChatGPT reassured him that it was real." (Hill & Freedman, New York Times, August 2025)
This guide was drafted with the machine. It got things wrong with a straight face. It told me a 2017 study said something the study never said, and it sounded exactly as sure as when it was right. I caught it because I opened the study. If it can't show you where it read something, you haven't been told anything yet.
Serious harm is rare. OpenAI's own early estimate is that about 0.07% of weekly users show possible signs of psychosis or mania. A review of nearly 54,000 psychiatric records in Denmark found 38 patients with possible harm linked to chatbot use, mostly worsening delusions, with no direct cause proven. (OpenAI, October 2025; Olsen, Reinecke-Tellefsen & Østergaard, Acta Psychiatrica Scandinavica, 2026)
The quieter risk is leaning on it. In a four-week trial with 981 people, the ones who chose to use the chatbot most ended up lonelier and more dependent on it. That's a correlation, and lonely people may simply use it more. Either way, it's worth knowing which way you're drifting. (Fang et al., MIT Media Lab and OpenAI, 2025)
"It was not genuine and felt like the ultimate masking."
That's an autistic user on running every email through ChatGPT. Among the risks named in the posts, about a third were exactly this: the tool slowly replacing a person's own voice. Another third clashed with the autistic sense of justice, over environmental cost, data, and creative work used without pay. (Ma et al., 2026)
The machine also brings its own ideas about us. In a 2025 study, GPT-3.5 talked up the importance of representation while "perpetuating mostly negative biases about autistic people." (Park et al., CHI 2025) A 2026 study of 345,000 answers from six AI models found that when users said they were autistic, the models disproportionately steered them away from social events, confrontation, new experiences and romantic relationships. One autistic participant's reaction became the paper's title, and it's the best line in this whole guide: "Are we writing an advice column for Spock here?" (Wohn et al., CHI 2026)
And in a 2025 survey of 200 autistic adults about negative self-talk, AI was the least helpful source of support they rated, below a therapist. (Carik et al., CHI 2025, as summarized)
Here's how I push back. I keep a folder of my own writing that goes back to 2004. When the machine drafts these guides, it has to sound like that folder, and there's a list of habits it isn't allowed to have. If it rewrites you into someone your friends wouldn't recognize, that's the mask. You don't need twenty years of blog posts to fight it. Three of your own messages will do. Paste them in and tell it to write like those.
Nothing in the research says stop. It says keep a human in the loop, and keep yourself in the words.
Draft in your own voice, then ask how it might land. Take the notes you agree with and leave the rest. Your phrasing is the part that's you.
Ask it for the strongest case that you're wrong. Agreement is its natural pull. Anything that feels like a revelation gets checked with an actual person before you act on it.
If the advice is "make eye contact and smile," or "maybe skip the party," notice it. The models push masking and steer autistic users away from people.
Breaking a task into pieces is one of the most common uses and one of the best. Let it hand you the first step, then go do the first step.
Heavier use went with more loneliness and dependence. If the chatbot has become the only one you talk to, that's worth noticing. Bring it here.
A rehearsal partner that doesn't flinch is useful. The rehearsal is for the real conversation. Rooms like this one exist so that conversation can be with a human who won't flinch either.
Paste in three messages you wrote and tell it to match them. If a rewrite stops sounding like you, ask for it back in your own words, or keep yours.
Ask for the source, then open it. A confident answer and a wrong answer sound exactly the same. If there's no source, treat it as a guess.
Answer one, or all of them, or just listen. Chat counts.
You don't have to say what it was. Just why it was easier.
The email it untangled, the meltdown it talked you down from, the task it broke into pieces.
Bad advice, the stereotype, the time it agreed with you and shouldn't have.
Be specific. That's a list the people in your life could actually use.
The autism and AI studies are small or drawn from social media, and they mostly hear from people who type easily. Several are preprints or conference papers. The guide says where a finding is early.
This one can bring up loneliness, and the sting of noticing that a program was kinder than the people around you. That's worth saying out loud, and it's exactly the kind of thing this room is for.
I use the machine to do the work of people we don't have yet. It drafts. It checks. It doesn't show up on Monday. You do.
Bring the thing you'd usually only type to a chatbot, if you want to, and see what a room full of people does with it.