AI Communication Coaching for Autistic Adults: The Research

Research on AI communication coaching for autistic adults is early and small. A 30-person Stanford randomized trial found an AI coach called Noora improved empathetic responses, a three-person follow-up saw gains in a workplace internship, and a CHI 2024 study found autistic workers preferred GPT-4 advice that a job coach questioned. Promising, not proven.

This guide walks through those studies one at a time: who took part, what was measured, what changed, and what the authors themselves say they did not test. It is written for autistic adults deciding whether these tools are worth their time, and for families, employers, and support staff trying to separate real findings from marketing.

What Is AI Communication Coaching?

AI communication coaching means using a large language model, the technology behind tools like ChatGPT and Claude, to practice or plan conversations. In research settings, that usually takes one of two forms. The first is a structured practice program: the AI presents a conversational situation, the person responds, and the AI gives feedback. The second is open-ended advice: a person describes a real situation, such as a confusing message from a coworker, and asks the AI what is going on or how to reply.

Both forms raise the same basic question. Does the practice or advice actually help the person in real conversations with real people, and on whose terms? The studies below give partial answers, and they do not all point in the same direction.

What Did the Stanford Noora Trial Find?

The most rigorous study so far is Using Artificial Intelligence to Improve Empathetic Statements in Autistic Adolescents and Adults: A Randomized Clinical Trial, by Lynn Kern Koegel and colleagues at the Stanford University School of Medicine and Stanford's Department of Computer Science. It was published online in the Journal of Autism and Developmental Disorders in February 2025.

The program, called Noora, presents a statement where an empathetic reply would fit, asks the participant to respond, and gives feedback on the answer. Here is how the trial was set up, according to the full paper on PubMed Central:

• Who took part. 30 autistic adolescents and adults, ages 11 to 35, who could hold a conversation but had difficulty with empathetic responses. 15 were randomly assigned to use Noora and 15 to a waitlist control group.
• How much practice. Participants were asked to respond to 10 statements a day, 5 days a week, for 4 weeks, for an expected total of 200 practice trials.
• How results were measured. Researchers collected conversation samples before and after the intervention, to see whether any change showed up in natural conversation, not only inside the app.

The headline result: in the post-intervention conversation samples, the Noora group's rate of appropriate empathetic responses rose from an average of 16.67 percent to 50.94 percent. The waitlist group moved from 28.87 percent to 31.40 percent. Inside the program itself, 10 of 14 participants analyzed (71 percent) showed an improving trend from their first week of practice to their last.

The authors describe this as evidence that an AI program can improve empathetic responses in a short time and that the gains carried over into conversation. That is a real finding from a randomized design, which is more than most AI-and-autism claims can point to.

Did the Gains Carry Over to Real Workplaces?

A follow-up study, Using Artificial Intelligence to Support Emotionally Responsive Verbal Communication Among Autistic Workers in a Work Internship Setting, was published online in the same journal in April 2026. Koegel worked with researchers from the Rehabilitation Research and Training Center at Virginia Commonwealth University to try Noora with autistic young adults in an intensive vocational internship program.

This was a single-case study with three participants, using what researchers call a concurrent multiple baseline design: each person started the program at a different time, so changes could be linked to when the AI practice began. The outcome was measured with human-delivered conversation probes built around workplace-relevant scenarios. According to the abstract, all three participants improved right after starting Noora, and the gains showed up in practice sessions, in the conversation probes, and at a later retention check.

Three people is a very small number. The study shows the program can be used inside a real job-training setting, which matters, but it cannot tell anyone how well it would work across a larger and more varied group of autistic workers.

Why Do Some Autistic Adults Prefer AI for Communication Advice?

A different line of research asks what autistic adults actually want from these tools. At the CHI 2024 conference, Carnegie Mellon researchers JiWoong Jang, Sanika Moharana, Patrick Carrington, and Andrew Begel presented "It's the only thing I can trust": Envisioning Large Language Model Use by Autistic Workers for Communication Assistance.

Ready to Get Started?

Tell us your biggest time-waster. We will map out your AI system.

Quote My Agent

They asked 11 autistic participants to bring real questions about workplace social difficulties to two sources: a GPT-4-based chatbot and a disguised human confederate. Participants strongly preferred the chatbot. As Carnegie Mellon's summary of the study put it, they valued the format of its answers, and one participant said it was "the only thing I trust because not every company or business is inclusive."

But a coach who specializes in supporting autistic job seekers reviewed the exchanges and raised concerns that the chatbot was giving questionable advice. In one example from the CMU summary, a participant asked how to make friends at work and the chatbot suggested simply walking up to people and starting a conversation, advice that ignores how uncomfortable that can be for many autistic people.

The researchers did not treat this as a simple case of "the AI was wrong." They framed it as a tension between what autistic users want and what practitioners consider good advice, and they recommended designing these tools to center autistic experiences. That framing matters, because both sides can be partly right.

A broader study backs up the picture. Exploring Large Language Models Through a Neurodivergent Lens, by Buse Carik, Kaike Ping, Xiaohan Ding, and Eugenia H. Rho of Virginia Tech, published in the Proceedings of the ACM on Human-Computer Interaction in 2025, analyzed discussions from 61 neurodivergent communities on Reddit between November 2022 and January 2024. It found 20 use cases across five areas, including interpersonal communication. The challenges people reported included overly neurotypical responses and the limits of text-only interaction, and their concerns included false information, overreliance, and AI replacing human connection. This study covers neurodivergent people broadly, not only autistic adults, so it adds context rather than autism-specific evidence.

What Are the Limits of the Research So Far?

Every study above is worth reading, and every one has limits that its authors or the evidence make clear:

• Small samples. The trials so far involve 30 people and 3 people. The CHI study had 11 participants. None of these is large enough to say how a tool works across the full range of autistic adults.
• No head-to-head comparison with human support. The Noora trial compared the AI to a waitlist group, not to a human-led program. The authors list this as a limitation in the paper.
• Mixed ages. The randomized trial included adolescents as young as 11 alongside adults up to 35, so its results are not specific to adults.
• Imperfect AI grading. In the trial's own reliability analysis, Noora's feedback was judged helpful 95 percent of the time, but its grading matched human raters 85 percent of the time. It sometimes marked a reasonable attempt at empathy as incorrect, often because of small grammar or spelling issues.
• Who sets the target. These programs measure progress against a particular idea of an appropriate response. In 2012, autistic scholar Damian Milton described the double empathy problem, the argument that misunderstandings between autistic and non-autistic people run in both directions. Training only one side of a conversation does not address the other side.
• Disclosures. The Noora trial paper discloses that its lead author is the editor of the journal that published it and was not involved in the editorial process, and that she is a partner in a private consulting company. Readers can weigh that for themselves.

None of this means the findings are wrong. It means the honest summary is that the evidence is early: one small randomized trial, one three-person field study, and a few qualitative studies of how people use general chatbots.

How Can Autistic Adults Use AI for Communication on Their Own Terms?

The research points toward a few practical habits. None of these are medical advice or therapy. They are ways to keep a general-purpose AI tool useful while staying in control of it.

• Use it to decode, then decide. Asking an AI what a vague email might mean, or for two or three ways to reply, can lower the effort of a confusing exchange. The choice of what to send stays with you.
• Tell it how you communicate. The Reddit study found people rewriting prompts to make responses less neurotypical. A short note at the top, such as "I prefer direct, literal answers with no idioms," often changes the tone of every reply.
• Check advice that involves other people. The CHI study showed that confident advice can still miss the mark. Treat social suggestions as options to test, not rules to follow, and run anything high-stakes past someone you trust.
• Keep the human connections that work. Participants in the Reddit study worried about AI replacing human connection. AI practice can sit alongside the people who already support you rather than replace them.
• Watch what you share. Anything you type into a chatbot goes to the company that runs it. Leave out medical details and other people's private information unless you are comfortable with that.

If you want to go further, our guide on how to build a Claude Skill for autism support shows how to save literal, autistic-affirming instructions so you do not have to retype them in every conversation. For a wider view of the tools available, see our roundup of the best AI tools for autistic people and their families.

What Should Families and Employers Take From This?

For families, the takeaway is to let the autistic person lead. The CHI study's participants valued AI because it felt safe and nonjudgmental, and that value disappears if the tool becomes another way to correct someone. Ask what kind of help they want, and respect the answer.

For employers, the workplace studies suggest that structured AI practice can fit inside a job-training program. They do not suggest that an app replaces an inclusive workplace. The CHI participant who called AI "the only thing I trust" was describing a gap in human support, and closing that gap is the employer's job, not the employee's.

For everyone, the honest position is cautious interest. The first randomized evidence is encouraging, the field is moving, and much larger studies, designed with autistic people rather than only about them, are what will settle the open questions.

Sources

Frequently Asked Questions

Does AI communication coaching work for autistic adults?

Early evidence is encouraging but limited. A 30-person Stanford randomized trial found the AI program Noora improved empathetic responses in conversation compared with a waitlist group, and a three-person workplace study saw similar gains. Larger studies are needed before anyone can call it proven.

What is Noora?

Noora is an AI program built by Stanford researchers that uses a large language model to coach conversation. It presents a statement where an empathetic reply fits, asks the person to respond, and gives feedback. It has been tested in a randomized trial and a small workplace internship study.

How many people were in the Noora trial?

The randomized trial enrolled 30 autistic adolescents and adults, ages 11 to 35, with 15 using Noora and 15 in a waitlist control group. Participants were asked to practice with 10 statements a day, 5 days a week, for 4 weeks.

Why do some autistic adults prefer AI chatbots for advice?

In a CHI 2024 study of 11 autistic workers, participants strongly preferred a GPT-4 chatbot over a disguised human. A Carnegie Mellon summary says they valued its answer format, and one participant said it was the only thing they trusted because not every workplace is inclusive.

Is AI communication advice always reliable?

No. In the same CHI 2024 study, a coach who supports autistic job seekers flagged some chatbot advice as questionable, such as suggesting someone simply walk up to coworkers to make friends. Treat AI social advice as options to consider, not instructions.

Is AI communication coaching a form of therapy?

No. The tools discussed here are practice and drafting aids, and nothing in this article is medical advice. The research is early, so decisions about clinical support belong with the autistic person and the professionals they choose to work with.

Build Your AI System

Tell us what is costing you the most time. We will map out exactly what your business needs. Free, no obligation.

Build An Agent