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

Paper Title

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

Publication Info

• Topic area: Ultra-personalized AI in Augmentative and Alternative Communication (AAC) systems.
• Keywords: AAC, ultra-personalization, large language models, agency, identity, privacy, autoethnography, assistive technology, AI-mediated communication.

Background and Problem

• Problem / challenge: Generic AI auto-complete systems fail to capture the personal identity of AAC users, requiring significant editing effort. While personalization is technically feasible, its impact on agency, identity, and privacy is not well understood.
• Significance: AAC users rely on such systems for everyday communication, making it crucial to balance fluency, expressivity, and user control. Missteps in personalization can undermine identity and privacy, with real-world consequences.
• Motivation and related work: Prior AAC systems have explored predictive text and context-aware suggestions but often lack personalization. Research highlights the need for systems that preserve user authorship and identity while addressing concerns about privacy and contextual appropriateness. This study builds on these insights by examining ultra-personalized AI trained on a single user’s communication history.

Solution

• Proposed approach: Development and evaluation of an ultra-personalized AAC system using a language model fine-tuned on the lead author’s communication data, explored through an autoethnographic study.
• Novelty:
  1. Empirical insights into how ultra-personalized AAC shapes user agency, identity, and privacy.
  2. Identification of trade-offs in personalization, including self-censorship and privacy risks.
  3. Use of lived experience to document real-world dynamics of AI-mediated communication over time.
• Procedure and key techniques:
  1. Phase 1: Seven months of data collection using a custom AAC app, logging all face-to-face communication.
  2. Phase 2: Data curation and fine-tuning a GPT-4.1-mini model on the collected dataset, with filtering for sensitive content.
  3. Phase 3: Three months of daily use of the personalized AAC system, with structured diary reflections and usage analytics.

Results

• Concrete findings:
  • The personalized model improved communication speed (mean WPM: 41.7 vs. baseline 31.4) and aligned with the user’s bilingual and cultural identity.
  • Suggestions were accepted selectively, with an average of 7.5 words per accepted suggestion.
  • Privacy concerns arose when the model resurfaced sensitive or contextually inappropriate details.
• Advantage over baselines:
  • Enhanced expressivity and fluency compared to generic models, with better alignment to the user’s tone and cultural-linguistic patterns.
  • Partial suggestion acceptance preserved user agency and authorship.
• Experiments / evaluation:
  • Usage analytics tracked metrics like word acceptance rates and suggestion alignment.
  • Diary reflections revealed themes around agency, identity, and privacy, highlighting both benefits and risks of ultra-personalization.
• Limitations and future work:
  • Single-user study limits generalizability; findings are most applicable to literate AAC users with text-entry devices.
  • Model performance was not optimized for state-of-the-art accuracy; future work should explore multi-user studies, controlled evaluations of latency, and adaptive systems.

Summary

This study explores the socio-technical implications of ultra-personalized AI in AAC through a single-user autoethnographic account. By fine-tuning a language model on the user’s communication data, the system enhanced fluency and expressivity while reflecting the user’s bilingual and cultural identity. However, challenges such as self-censorship, contextual missteps, and privacy risks emerged. The findings emphasize the need for AAC systems that balance agency, identity, and privacy, with modular, adaptive designs that evolve alongside users. This work provides early insights into the potential and risks of ultra-personalized AAC, offering design directions for future implementations.