TwIPS: A Large Language Model Powered Texting Application to Simplify Conversational Nuances for Autistic Users
Rukhshan Haroon, Fahad Dogar · 2024 · ASSETS '24: Proceedings of the 26th International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3663548.3675633
Summary
This paper presents TwIPS, a prototype texting application powered by a large language model that assists autistic users with the pragmatic and tonal aspects of text-based communication. Many autistic individuals experience difficulties interpreting non-literal language (sarcasm, irony, implied meanings), conveying intended emotional tone in their own messages, and navigating social expectations around communication style. Many also engage in "masking" — consciously adjusting their communication to appear more neurotypical — which requires significant cognitive effort and emotional labour. TwIPS addresses these challenges through three core features: a Decipher function that explains the likely tone and meaning of incoming messages (e.g., identifying sarcasm, passive aggression, or genuine friendliness); a Tone Check function that analyses the user's draft message and indicates what emotional tone it might convey to the recipient, flagging potential mismatches with the user's intent; and a Rephrase function that suggests alternative phrasings when a message might be misconstrued or received negatively. The system is built on GPT-4 with carefully designed prompts that provide conversational context. The evaluation involved 8 autistic participants in an in-lab study using an AI-based simulation — rather than texting with real people, participants interacted with an AI conversational partner following scripted scenarios covering common challenging situations (receiving ambiguous messages, navigating conflict, responding to sarcasm, managing emotionally charged conversations). Post-task interviews explored participants' experiences, preferences, and concerns about AI-mediated communication.
Key findings
Participants found TwIPS useful across all three functions, with the Decipher feature being most valued — participants described the relief of having an on-demand "translator" for ambiguous messages rather than spending extended time and mental energy trying to parse social cues. The Tone Check was appreciated for providing a "second opinion" before sending messages, helping participants catch unintentionally blunt or harsh phrasings. Several participants noted that TwIPS offered a better alternative to tone indicators (e.g., /s for sarcasm) because it worked with any message rather than requiring the sender to add markers. A key finding was that TwIPS facilitated constructive self-reflection on writing style — participants reported learning about how their communication patterns might be perceived, which some described as more valuable than the immediate assistance. However, the study revealed important tensions around autonomy and authenticity. Some participants worried about over-reliance on AI mediation, fearing it could erode their own communication skills or make their messages feel inauthentic ("it's not me anymore"). The masking concern was particularly nuanced: while TwIPS could reduce the cognitive effort of masking, some participants questioned whether the tool was essentially automating masking rather than promoting acceptance of autistic communication styles. Trust calibration was a recurring theme — participants needed to develop appropriate trust in the AI's tone interpretations, and several noted cases where the AI's assessment of tone was wrong or culturally biased. Customisation was strongly desired: participants wanted to adjust the AI's sensitivity level, specify their personal communication norms (e.g., "I use direct language and that's okay"), and control when features activated.
Relevance
TwIPS addresses a genuine and underserved accessibility need — the pragmatic language challenges that many autistic people face in text-based communication, which is increasingly the dominant mode of social and professional interaction. For accessibility practitioners, the work raises important questions about the boundaries of AI-assisted communication: when does assistive support become automated masking, and how should tools balance helping users navigate neurotypical social expectations while respecting autistic communication as valid in its own right? The finding that self-reflection was valued as much as immediate assistance suggests that the most effective communication tools may be those that build understanding rather than just providing answers. The customisation needs expressed by participants — controlling sensitivity, defining personal norms, choosing when to activate features — reflect the broader principle that neurodivergent users need agency over how assistive AI interacts with their communication, not one-size-fits-all mediation. The tension between reducing masking effort and automating masking is a critical ethical consideration for any AI tool designed for neurodivergent users and deserves further exploration in the field.
Tags: autism · communication · large language models · text messaging · tone interpretation · masking · social cognition · AI-assisted communication · neurodiversity