Reimagining Support: Exploring Autistic Individuals' Visions for AI in Coping with Negative Self-Talk

Research Background and Issues

• What problems or challenges did the authors identify?
  The phenomenon of "negative self-talk" (NST) is prevalent among individuals with autism, leading to an increased risk of anxiety and depression. However, traditional therapeutic approaches, such as Cognitive Behavioral Therapy (CBT) or Acceptance and Commitment Therapy (ACT), often have limited effectiveness due to communication difficulties in some autistic individuals or constraints related to cost and resources. Meanwhile, large language models (LLMs) like ChatGPT and Claude have emerged as accessible support tools, prompting many individuals with autism to explore their use. However, these tools face inherent challenges, including "neurotypical biases" and suboptimal response quality.

• Why is this issue important?
  The issue of NST in the autistic population not only impacts their mental health but also undermines their performance in work, social relationships, and other areas. While AI tools have the potential to provide support, the unique needs of this population are not adequately understood or addressed by current models. Therefore, exploring how to optimize AI assistance to meet the needs of autistic users is of critical importance.

• Research Motivation and Related Work
  Existing research has explored the role of AI in mental health support but lacks in-depth analysis specific to the autistic population. Due to the diverse preferences and cognitive styles of autistic users, current LLMs often fail to fully understand and support their needs. Additionally, the research motivation includes examining whether autistic individuals and mental health practitioners would accept LLM-based mental health support mechanisms.

Solutions

• What methods or solutions did the authors propose?
  The authors conducted an online survey and interviews with mental health practitioners to explore the aspirations of autistic users for AI support and the potential of AI in managing NST. The study surveyed 200 autistic adults and analyzed their needs for different types of AI support, interaction styles, and content, incorporating professional opinions from mental health practitioners. Additionally, the authors evaluated the responses of existing LLMs, such as ChatGPT and Claude.

• What is innovative about this solution?
  The study not only comprehensively captured the diverse needs of the autistic population but also combined clinical evaluations from mental health practitioners regarding the response patterns of existing LLMs. This clarified the potential and limitations of LLMs in assisting autistic users. Notably, the authors considered multimodal interactions (e.g., music, tactile feedback) to enhance AI adaptability for autistic individuals, a domain that has not been widely explored in existing research.

• What are the implementation steps and key technologies used?

  1. Survey: Consisting of six sections, including assessments of NST frequency and impact, autistic users' aspirations for AI support, preferences for AI interaction, and concerns about using AI.
  2. Interviews: Collecting mental health practitioners' perspectives on the potential and challenges of AI-assisted mental health support.
  3. LLM Evaluation: Using hypothetical NST-related dialogue scripts designed by participants to generate AI responses (ChatGPT and Claude), which were then evaluated by mental health practitioners for professionalism and suitability.
  4. Data Analysis: Conducting statistical analysis on quantitative survey data, coding qualitative responses, and performing thematic analysis on LLM outputs.

Research Findings

• What specific findings were obtained?

  1. Characteristics and Impact of NST: NST revolves around negative self-perceptions, social evaluation, and social isolation. Triggers for these negative thought cycles include work stress, routine disruptions, or difficulties in social interactions.
  2. AI Aspirations of the Autistic Population: Participants hoped AI could assist in managing NST by identifying and restructuring negative thoughts, providing meditation guidance, and facilitating journaling.
  3. Feedback from Mental Health Practitioners: Current LLM responses often include vague or overly general expressions and excessive use of "technical jargon," which confuses users.
  4. AI Interaction Preferences: Autistic users expressed a high demand for multimodal interactions (e.g., music and visual responses) and interaction styles tailored to individual preferences. Additionally, humanized and gradual emotional tones were seen as safe and effective design directions.

• What advantages does it have compared to existing solutions?
  This study proposed design directions that address real-world user scenarios and clinical feedback, rather than focusing solely on technical functionality. By suggesting multimodal support and emphasizing personalization and neurodiversity in design principles, the research significantly surpasses previous discussions limited to single-text-based interactions.

• What were the experimental or evaluation results?
  Experimental data indicated that current LLM responses to NST interactions for autistic users remain suboptimal, particularly due to limited understanding of their specific needs and overly verbose language. Additionally, the neurotypical biases of AI models were identified as a significant issue, which, if unresolved, could further erode trust among the autistic population.

• Limitations and Future Directions
  Limitations:

  • The sample primarily consisted of participants from the United States, predominantly white and highly educated, lacking diversity in racial and cultural backgrounds.
  • The experimental scope was limited to ChatGPT and Claude, excluding non-commercial or highly customized generative models.
  • The sample size of mental health practitioners was small (only three participants), and did not cover all major therapeutic domains.

  Future Directions:

  • Conduct cross-cultural studies to capture user needs in other geographic regions.
  • Introduce a broader range of customized AI models and explore new multimodal interaction solutions.
  • Expand the scope of interviews with mental health practitioners to include perspectives from other therapeutic approaches (e.g., art or music therapy).
  • Develop more transparent and trustworthy AI systems to address current concerns about data privacy and reliability.

This paper thoroughly explores the potential of AI in supporting the mental health of the autistic population and proposes diverse design recommendations, providing a solid foundation for future technological development and academic research.