“It’s the only thing I can trust”: Envisioning Large Language Model Use by Autistic Workers for Communication Assistance

Document Title

“It’s the only thing I can trust”: Envisioning Large Language Model Use by Autistic Workers for Communication Assistance

Document Information

• Research Domain: Human-Computer Interaction and assistive technologies for autistic individuals
• Keywords: Autism, social communication, Large Language Models (LLM), workplace, artificial intelligence, neurodiversity, assistive technology, design insights, work environment, workplace adaptation

Research Background and Problem Statement

• Identified Challenges:

  1. Autistic adults often face communication difficulties in the workplace, including understanding implicit social norms of non-autistic individuals, handling ambiguous or unclear information and expectations, and navigating these social challenges without adequate resources.
  2. Many autistic individuals, due to prevalent workplace discrimination and social complexities, lack sufficient guidance and support, turning to family or friends for help, which can lead to emotional harm from conflicting opinions.
  3. Autism-specific workplace resources (e.g., job coaches or vocational rehabilitation specialists) are often difficult to access, forcing individuals to seek alternative forms of support.

• Research Significance:

  1. Large Language Models (LLMs) have recently demonstrated powerful language understanding and generation capabilities, offering potential for interpreting social situations and improving communication.
  2. If this technology can be utilized by autistic individuals to alleviate social communication challenges, it could address significant gaps in existing resources and services.

• Research Motivation: The researchers aim to explore whether LLMs can serve as effective assistive tools for autistic adults in addressing social communication challenges in the workplace, while evaluating potential risks and directions for design improvements.

Solution

• Research Methods:

  1. The authors organized interactions with two chatbots involving 11 autistic participants: one based on OpenAI GPT-4 (referred to as Paprika) and another human-mimicking “chat agent” (referred to as Pepper).
  2. Participants discussed real workplace social issues with both chatbots and evaluated their performance.
  3. A professional career counselor was invited to assess the quality of the chatbots’ responses.

• Methodological Innovations:

  1. A comparative experiment was designed, incorporating both automated LLM interactions and human-like “real person” interactions (disguised dialogue).
  2. Dual data collection pathways were implemented, combining participants’ quantitative ratings, qualitative descriptions, and expert analysis of LLM-generated content.

• Implementation Steps and Key Technologies:

  1. Customized "prompt engineering" was applied to the LLM to ensure the generated content adhered to formal workplace communication styles.
  2. All experiments were conducted via an online platform (Discord), and researchers performed structured coding analysis on chat records and interview content.
  3. Cross-validation of experimental findings was conducted using two types of participant data: subjective preferences and cognitive analysis of content quality.

Research Findings

• Summary of Results:

  1. 82% of participants preferred using LLMs as social communication tools over the human-mimicking “chat agent.”
  2. Participants generally found LLM responses to be clearer and more structured, offering “unbiased, detailed, and friendly” advice compared to the human agent.
  3. Many participants reported that using LLMs enabled them to independently explore solutions to workplace social issues without the emotional risks associated with seeking help from real people.

• Advantages Over Existing Resources:

  1. Convenience: LLMs provide immediate, barrier-free communication guidance, far surpassing traditional resources that are only available in specific times and settings.
  2. Privacy Protection: Many participants noted that interacting with LLMs reduced the stress of being judged for disclosing problems or identity.
  3. Cost-Effectiveness: Compared to hiring workplace coaches, LLMs offer significantly higher affordability.

• Challenges and Limitations:

  1. Professional counselors pointed out that LLM suggestions often reflect “neurotypical” biases, potentially misleading autistic individuals to mimic non-autistic social norms (e.g., forced eye contact or specific body language), thereby increasing psychological burden.
  2. Some LLM recommendations were overly optimistic, failing to account for potential risks of failure or conflict in real-world scenarios.

• Future Research Directions:

  1. Optimize LLM training datasets to include more diverse autistic user perspectives, enhancing recognition of unique needs among autistic users.
  2. Investigate effective integration of Value Sensitive Design (VSD) to ensure the technology responds to user needs rather than reinforcing existing structural barriers in the workplace.
  3. Deepen research into the “implicit social impacts” of LLMs, such as how widespread use of these tools might further shape autistic individuals’ social identities and workplace rights.