This paper, presented at CHI EA ’26, introduces NeuroWise, an AI-powered practice environment designed to bridge the communication gap between neurotypical (NT) and autistic individuals. By moving away from the traditional model of "training" autistic people to fit neurotypical norms, NeuroWise focuses on helping neurotypical partners understand and adapt to autistic communication styles.
TL;DR
NeuroWise is a multi-agent LLM system that coaches neurotypical users during simulated interactions with an autistic partner. By visualizing the partner's stress and interpreting their internal experiences in real-time, the system successfully prevents users from blaming "deficits" and helps them reach faster, more empathetic resolutions.
What Problem Does This Paper Address?
The authors ground their work in the "Double Empathy Problem," which suggests that communication breakdowns between neurotypical and autistic people are a result of mutual misunderstanding, not just a "deficit" in the autistic person.
However, most existing interventions are one-sided, focusing on teaching autistic individuals how to "mask" or adapt to neurotypical expectations. Neurotypical individuals often struggle to interpret autistic social cues, and when interactions become difficult, they tend to fall back on deficit-based attribution—blaming the breakdown on the other person's autism rather than seeing it as a situational mismatch. There is a distinct lack of tools that help neurotypical people practice perspective-taking in a safe, guided environment.
Methods
NeuroWise is a web-based platform (built with Streamlit and OpenAI’s GPT-4o-mini) that features a multi-agent architecture. When a user chats with "Alex" (the simulated autistic partner), three specific agents provide "glass-box" transparency:
- Stress Estimator: A hybrid algorithm that classifies user messages (e.g., validation vs. pressure) and updates a visual 0–100% stress bar for the partner.
- Interpreter Agent: Triggered by stress spikes, this agent explains why Alex might be feeling stressed, making "invisible" sensory or cognitive states (like the overwhelming smell of food or a disrupted routine) visible to the user.
- Coach Agent: Provides contextual suggestions for how to respond, focusing on accommodation and validation rather than pressuring the partner to comply.
The researchers evaluated the system through a between-subjects study with 30 neurotypical adults, comparing NeuroWise to a baseline chatbot without the coaching features.
Datasets
The study did not rely on a large-scale pre-existing dataset for training. Instead:
- Validation Set: The Stress Bar algorithm was pre-validated against a custom dataset of 15 scripted conversations (63 turns) rated by two independent human experts.
- Evaluation Scenario: Participants interacted with a custom-designed scenario involving a "Friday pizza night" routine that is disrupted by the user bringing home Thai food—a situation designed to trigger common autistic challenges like sensory sensitivity and routine dependence.
Key Insights and Ideas
- The "Regression" Risk: One of the most striking findings was that in the baseline group (no AI support), difficult interactions actually made people less empathetic. These users shifted toward blaming autistic "deficits" after the conversation.
- Designing for Attribution, Not Just Behavior: The authors argue that simply telling a user what to say (the "Coach") is less effective than explaining why the partner is reacting a certain way (the "Interpreter"). When users understand the root cause, they naturally adjust their behavior.
- The Power of Visualization: Making "invisible" internal states (like stress) visible through a UI element (the Stress Bar) allows users to see the immediate impact of their communication style without feeling like they are being "corrected" by a teacher.
Key Results
- Reduced Blame: NeuroWise users showed a significant reduction in deficit-based framing ($p=0.02$), while baseline users increased their endorsement of deficit-based explanations.
- Increased Efficiency: Users with NeuroWise completed the conversation task 37% more efficiently (fewer turns) than the baseline group ($p=0.03$), without sacrificing the quality of the resolution.
- High Acceptance: Participants rated the system as highly helpful (6.6/7), with the Interpreter being the most valued feature. Users noted that it helped them "actually know how my partner reacts."
Limitations and Future Work
The authors acknowledge several limitations:
- Small Scale: The study involved only 30 participants and a single simulated scenario.
- Simulated vs. Real: The effects were measured in a controlled, simulated environment; it remains to be seen if these attributional changes transfer to real-world, long-term relationships.
- One-Way Support: The current tool is unidirectional (NT learning about ND). The authors suggest that future "bidirectional" tools could help both parties interpret each other's patterns simultaneously.
- Diversity of the Spectrum: The simulated partner "Alex" represents a specific set of traits; future versions should be personalizable to reflect the broad diversity of the autism spectrum.
Summary updated