I, Robot? Exploring Ultra-Personalized AI-Powered AAC; an Autoethnographic Account

TL;DR

This study examines ultra-personalized AAC by autoethnographically training a user-specific predictive system on the sole user’s speech history and evaluating it across three phases: long-term data collection, data curation with targeted fine-tuning, and three months of daily use. It finds that personalization can both enhance fluency and expressivity and reshape agency, identity, and privacy, sometimes resurfacing intimate or context-inappropriate details. The work argues for context-aware surfacing, modular identity components, and user-controlled tuning to preserve authentic voice while safeguarding privacy and social relations. Its insights aim to guide practical, ethically aware deployment of ultra-personalized AAC tools that adapt to users over time.

Abstract

Generic AI auto-complete for message composition often fails to capture the nuance of personal identity, requiring significant editing. While harmless in low-stakes settings, for users of Augmentative and Alternative Communication (AAC) devices, who rely on such systems for everyday communication, this editing burden is particularly acute. Intuitively, the need for edits would be lower if language models were personalized to the communication of the specific user. While personalization has been shown to be technically feasible, it raises questions about how such systems affect AAC users' agency, identity, and privacy. To understand how these shifts in practice, we conduct an autoethnographic study in three phases: (1) seven months of collecting all the lead author's AAC communication data, (2) fine-tuning a model on this dataset, and (3) three months of daily use of personalized AI suggestions. Observations across these phases include that logging everyday conversations reshaped the author's sense of agency, the model training selectively amplified or muted aspects of their identity, and suggestions occasionally resurfaced private details outside their original context. Our findings show that ultra-personalized AAC reshapes communication by continually renegotiating agency, identity, and privacy between user and model. We highlight design directions for building context-adaptive, user-controlled personalization AAC technology that supports expressive, authentic communication.

Figures (7)

• Figure 1: Shows the mode of conversational interaction: a) The lead author is typing a message. b) The lead author shows the message to his conversational partner.
• Figure 2: Shows the workflow of using the app. a) landing screen when you open the app, which shows the conversation threads where you can create a new one or select a previous one. b) an active thread, shows the title and context description (editable) and past messages, and you can create new messages to the active thread. c) composer modal, where the users would write and display the text, with controls for font size on top. The clear button would append the message to the active thread and clear the text, speak button reads the text out loud.
• Figure 3: Distribution of word count in messages.
• Figure 4: OpenAI Platform training accuracy curve (smoothing: 0.85).
• Figure 5: Example of text completion from the LLM