New research published in the Journal of Autism and Developmental Disorders provides evidence that artificial intelligence can help autistic individuals improve their conversational skills. Two recent studies tested a chatbot designed to guide users in making empathetic responses during social interactions. The findings suggest that short, regular practice sessions with this technology can lead to better communication in natural, face-to-face conversations.
Autism spectrum disorder is a developmental condition that affects how people communicate, learn, and interact with others. One common characteristic involves differences in social communication, such as knowing how to respond when someone shares an emotional experience. This specific conversational skill is often referred to as verbal empathy.
Empathy is generally understood to have two distinct parts. The cognitive component involves understanding the actual words and the intent behind what someone is saying. The affective component involves recognizing the underlying emotion, such as happiness, frustration, or sadness. Following the recognition of these two components, a person must formulate a verbal response that matches the situation to maintain a smooth conversation.
Challenges with verbal empathy can make it difficult to form friendships or participate fully in community activities. In the workplace, these communication differences can lead to misunderstandings with coworkers or supervisors. Because of these social barriers, employment success tends to be lower for autistic adults, who face disproportionately high rates of job loss and underemployment.
Traditional support programs often rely on face-to-face sessions with trained professionals to teach these social skills. These human-led programs are effective, but they can be expensive and hard to access for many families. There is also a severe lack of programs specifically designed for autistic adolescents and adults, leaving older individuals without adequate support.
“There are so many people on the autism spectrum that lack quality services or don’t receive services at all,” said Lynn Kern Koegel, a clinical educator at the Stanford School of Medicine. “For those who do, the face-to-face services can be very costly. We truly believe that AI can be a helpful and cost-efficient resource providing flexibility to practice at any time and location for improving verbal communication.”
Koegel helped initiate the project following an interaction at an event. “I was giving a speech at the Stanford Autism Conference and one of the attendees, who was in tech, approached me after and said that the face-to-face researched interventions I discussed could be supported through AI,” Koegel explained. She noted that they developed a prototype to offer support and feedback to autistic individuals facing conversational challenges.
Recent advances in artificial intelligence provided the necessary tools to scale this idea. “We were able to turn our prototype into reality when LLMs became more viable, so that users could practice responding and get feedback in real time,” Koegel said. Large language models power modern chatbots and can simulate highly realistic human conversation.
The project quickly grew into a large, interdisciplinary collaboration. “Eventually Monica Lam, Professor of Computer Science at Stanford, and her team along with my team of speech pathologists, psychologists, and special educators from the Stanford School of Medicine, teamed up to develop the programs,” Koegel told PsyPost. Together, they created a specialized artificial intelligence tool named Noora.
This program was designed to give autistic users a safe, self-paced environment to practice responding to emotional statements. The researchers wanted to see if practicing with Noora would translate into better communication with real humans. They designed the software to provide immediate feedback on users’ replies because human conversations are naturally unpredictable.
The team took specific steps to ensure the software was safe for the participants to use. They programmed the chatbot with pre-written scenarios to prevent it from generating false or inappropriate information. “Noora was not designed to be a ‘friend’ replacement, and by design, is not a multi-turn open domain conversation agent,” Koegel noted.
Instead, the program maintains a tight focus on skill-building. “It was designed with precise focus on very specific areas that may interfere with relationships, social connections, and employment,” Koegel said. She added that all the leading statements and examples of responses were verified by humans and tested by both autistic and neurotypical individuals.
In the first study, researchers conducted a randomized clinical trial with 30 autistic adolescents and adults between the ages of 11 and 35. Half of the participants were randomly assigned to use the Noora program immediately. The other half were placed on a waitlist to serve as a comparison group.
To participate, individuals had to be formally diagnosed with autism and be able to engage in a back-and-forth conversation for twenty minutes. They also had to demonstrate some difficulty with responding empathetically during a baseline screening test. The experimental group used the artificial intelligence program for about four weeks.
Participants in the experimental group were asked to complete 10 conversational practice trials per day, five days a week. During a session, Noora would present a leading statement, such as a complaint about a stressful job. The user would first rate whether the statement was positive, negative, or neutral before providing a spoken or typed reply.
The artificial intelligence analyzed the reply and provided immediate feedback on whether the response showed appropriate empathy. If a user provided a supportive response, the screen displayed celebratory confetti graphics. If a reply was off-topic or lacked support, Noora gently suggested a better alternative.
To measure progress, the researchers recorded a 20-minute video conversation between each participant and a human partner. These assessments occurred before and after the four-week period to track generalization, which means applying a learned skill to a new situation. Independent raters then scored these interactions to see if the computer practice made a difference.
The researchers found that practicing with the artificial intelligence translated to real-world human interactions. Before the intervention, the experimental group provided appropriate empathetic responses about 16.67 percent of the time. After four weeks of using Noora, their average score jumped to 50.94 percent, while the waitlist group showed almost no change.
“We were surprised at how quickly our adolescents and adults demonstrated improvements in the challenging areas and generalized their gains to human conversation,” Koegel said.
The study also tracked how well participants performed within the application itself. About 71 percent of the users demonstrated an improving trend from their first week of practice to their last week.
Surveys administered after the trial showed that 67 percent of participants completely enjoyed or enjoyed using the program. “Many people would like to improve their social communication, whether or not they have autism,” Koegel said. “Our ‘Noora’ program provides an opportunity to practice areas that may be interfering with successful social conversation.”
The second study expanded on these findings by testing the Noora program in an actual workplace setting. Getting and keeping a job can be challenging for autistic adults, and informal small talk is often necessary for workplace acceptance. The scientists focused on three autistic young adults, aged 19 to 22, who were participating in a hospital internship program.
In competitive employment environments, employees are expected to interpret social cues and adapt their communication accordingly. Being able to chat smoothly during a break or transition period helps build trust and collegiality among coworkers. An inability to navigate these unwritten social rules can lead to poor performance reviews or even termination.
The researchers used a multiple baseline design for this second experiment, staggering the start of the artificial intelligence training for each participant over time. This specific research method helps prove that the technology is responsible for the observed changes, rather than external factors. The three interns used the Noora program before or after their shifts, three days a week.
Throughout the study, job coaches engaged the participants in brief, unprompted conversations to test their skills in a natural work environment. The job coaches provided workplace-relevant statements, such as mentioning a mistake at work or feeling overwhelmed. The job coaches then recorded how the interns responded to these emotional cues over the phone or in person.
The data showed that all three participants experienced immediate improvements in their conversational skills right after they began using the software. For instance, the first participant initially responded with empathy during workplace tests only about 50 percent of the time. After starting the intervention, his scores consistently ranged from 75 to 100 percent.
The second participant showed a similar upward trend, moving from an average baseline of about 29 percent to steady, high scores. The third participant provided mostly brief, literal responses prior to the training. He showed significant improvement once he was allowed to use the program’s voice-to-text feature, eventually reaching 90 percent accuracy during the human conversation tests.
It is important to note that these tools are not intended to cure or change autistic individuals. The neurodiversity movement emphasizes that autistic communication differences are entirely valid. The Noora program was designed specifically for people who voluntarily requested support in navigating social expectations.
While these findings are promising, Koegel warned against using untested commercial products. “There are many apps and programs available, but most have not been scientifically validated, and gains beyond use in the actual programs are infrequently measured,” she explained. “Some of these programs may be helpful and others harmful. Our work stresses the importance of scientifically validating the effectiveness of AI programs for this population.”
The current software is also expanding to address other conversational skills. “Our first studies targeted verbal empathetic responses, but we also have a variety of additional modules targeting question asking, talking the right amount, giving compliments, and many other areas that can be challenging,” Koegel said.
There are some limitations to these studies that provide directions for future research. The first study compared the artificial intelligence program to a waitlist group rather than an active, human-led therapy group. The second study involved a very small sample size of only three individuals, meaning larger trials are needed to explore how these tools impact long-term career outcomes.
Koegel and her colleagues are now expanding their research to support younger age groups. “Our team, including computer scientist Tommy Bruzzese, has now developed an app for young children on the autism spectrum who are just starting to talk based on Pivotal Response Treatment (PRT), an evidenced based approach for improving communication, developed by Drs. Robert and Lynn Koegel,” she said. PRT is an established behavioral intervention that focuses on core areas of development, such as motivation.
“Teaching first words and communication takes a tremendous effort by parents and clinicians and can require support throughout the day,” Koegel noted. “To assist, our app teaches new words using short and contingent video rewards. We are currently researching how this active screentime, in place of stagnant screentime, may be helpful for parents and their children.”
The researchers are currently enrolling participants for this new study. Eligible families will receive free access to the program and can contact Elizabeth Ponder at lizziep@stanford.edu to learn more. Additional information about the team’s ongoing projects can be found at heynoora.com.
The study, “Using Artificial Intelligence to Improve Empathetic Statements in Autistic Adolescents and Adults: A Randomized Clinical Trial,” was authored by Lynn Kern Koegel, Elizabeth Ponder, Tommy Bruzzese, Mason Wang, Sina J. Semnani, Nathan Chi, Brittany L. Koegel, Tzu Yuan Lin, Ankush Swarnakar, and Monica S. Lam.
The study, “Using Artificial Intelligence to Support Emotionally Responsive Verbal Communication Among Autistic Workers in a Work Internship Setting,” was authored by Lynn K. Koegel, Paul Wehman, Christopher M. Claude, Monica S. Lam, Tommy Bruzzese, Melanie Derry, Alissa Brooke, Whitney Ham, Elizabeth Ponder, Trylanda J. Roane, Benjamin Rooney, and John Anderson.
Leave a comment
You must be logged in to post a comment.