People with autism have brains that are wired differently. This can make them especially strong in some areas—such as noticing patterns, remembering details, or thinking logically—while making other things like social cues or changes in routine more challenging.
There can also be stark differences in the way autistic and neurotypical people communicate, to the point where it may seem like each is using a different language, creating complications from social situations to the workplace.
For example, while non-autistic people often depend on nonverbal cues like body language and tone of voice, inferring emotion and intent, some autistic people rely on them less, and might interpret linguistic devices like sarcasm or irony literally.
Likewise, autistic people might prefer direct and clear communication—treating an indirect request (“When you get a chance, can you send that file?”) as genuinely non-urgent, interpreting a hedged refusal (“That might be difficult”) as uncertainty rather than a “no,” or taking a figurative expression (“This idea has legs”) literally.
On the other side, neurotypical people might misunderstand an autistic person’s direct and literal style as being blunt or unempathetic.
A team of Tufts scientists recently took up the challenge of creating a tool to bridge this communication gap. Instead of pushing autistic people to communicate in non-autistic ways, which can make social interactions inauthentic and cognitively draining for them—and which is the focus of many existing interventions—they created NeuroBridge, an AI-based learning tool that uses large language models to help neurotypical people learn how to better communicate with autistic people.
The researchers describe it in a new research paper published in the 27th International ACM SIGACCESS Conference on Computers and Accessibility, where it received the Best Student Paper Award.
“We received positive feedback on the utility of the application. Many neurotypical users were surprised to find the interpretations of the response options were obvious in hindsight, but never occurred to them.” Rukhshan Haroon
“NeuroBridge is not so much a tool to use on-demand to assist during interactions, like you might use a translator when traveling to a country with a different language,” said Rukhshan Haroon, a PhD candidate in the Department of Computer Science, who led the research project.
“It is more useful as a way for non-autistic people to gain firsthand experience with cross-neurotype communication, learn about autistic communication preferences, and use that understanding to adjust their own communication when interacting with autistic people,” he said.
“Through NeuroBridge,” he added, “our aim is to create an environment—among friends, co-workers, and organizations—that enables people to better recognize and appreciate neurodiverse communication styles, as well as the interdependent nature of social interactions.”
Fahad Dogar, an associate professor in the Department of Computer Science and at Tisch College for Civic Life, oversaw the project and said their approach was “grounded in the social model of disability, which emphasizes that disability arises not from individual deficits, but from the mismatch between individuals and their social environment.”
He noted that the system was developed with iterative feedback from a board of autistic volunteers, who helped improve its design and accuracy.
“We’re excited to build on this work and believe it has the potential for meaningful social impact,” he said. “We are already exploring ways to use it to enhance support for neurodiverse students at Tufts, collaborate with departments and campus resources that could benefit from it—such as the StAAR Center, which provides academic and accessibility support to students with disabilities—and pursue new opportunities to scale and evaluate its impact.”
How It Works
NeuroBridge creates a conversational scenario tailored for the user based on information that they provide about themselves, making it interesting and relatable.
At different points in the conversation, NeuroBridge presents the neurotypical user with three response options, each similar in meaning but varying in tone, clarity, or phrasing. For example, the user may ask it, “How can I speed up shoveling snow from my driveway?”
NeuroBridge then may present three different ways to phrase that question: Is there a way to speed up shoveling a driveway? Do you know how to speed up shoveling snow from a driveway? What methods can be used to speed up shoveling snow from a driveway?
It will point out that two of these options (those starting with ‘Is there a way…?’ and ‘Do you know…?’) can be interpreted differently than intended because they can be answered with ‘yes’ or ‘no,’ rather than advice on shoveling. The third option is clearest, because it explicitly asks for the information being sought.
The researchers note that the application tends to train users toward principles called Gricean maxims, developed by philosopher H. Paul Grice, that guide a conversational style that is clear, brief, orderly, and avoids ambiguity.
“We tested NeuroBridge with 12 individuals,” said Haroon. “We received positive feedback on the utility of the application. Many neurotypical users were surprised to find the interpretations of the response options were obvious in hindsight, but never occurred to them.”
The participants also found that the feedback the program provided helped them understand exactly what parts of their conversation could be received differently by an autistic person, making it useful for navigating future real-world interactions.
