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

1. Problem Statement

Augmentative and Alternative Communication (AAC) devices are essential for individuals with speech impairments, yet current generic AI auto-complete systems often fail to capture the nuances of a user's personal identity, tone, and cultural context. This forces AAC users to spend significant effort editing AI suggestions, creating a heavy cognitive and physical burden. While Large Language Models (LLMs) offer the technical feasibility of "ultra-personalization" (training on a specific user's data), there is a critical lack of empirical understanding regarding how such systems impact user agency (control over speech), identity (representation of self), and privacy (contextual boundaries) in real-world, long-term use.

2. Methodology

The authors conducted a three-phase autoethnographic study over a period of approximately 10 months, led by the first author (Tobias Weinberg), who is a full-time AAC user with a speech impairment.

• Phase 1: Data Collection (7 Months)

  • Setup: The lead author used a custom-built iOS app to log all face-to-face conversations.
  • Data Volume: 247 conversation threads containing 19,764 messages.
  • Observation: The author maintained reflective diary logs. A key finding was that the act of logging induced self-censorship; the author avoided profanity, dark humor, and sensitive topics to prevent the future model from learning them, thereby constraining his own agency during data collection.
  • Segmentation: Due to the difficulty of real-time conversation segmentation, data was organized by day rather than by discrete conversation threads.

• Phase 2: Curation and Model Training

  • Filtering: The dataset was curated to remove toxic content and utterances the author did not wish the AI to reproduce (e.g., profanity, specific jokes). This created a "socially well-behaved" model that reflected only the "good" parts of the author's identity.
  • Model Selection: Several LLMs (Gemma, Llama, Cohere, GPT-4.1-mini) were evaluated. GPT-4.1-mini was selected for fine-tuning because it best captured the author's bilingual (English/Spanish) style and cultural nuances (e.g., Argentine slang).
  • Training: The model was fine-tuned using Supervised Fine-Tuning (SFT) on OpenAI's platform. The training data was structured using an incremental chunking approach (predicting the next token based on partial input) to overfit the model to the author's specific patterns.
  • System Prompt: A custom system prompt was engineered to instruct the model to act as a personalized autocomplete for "Tobi," matching his tone and vocabulary without introducing new ideas.

• Phase 3: Deployment and Evaluation (3 Months)

  • Interface: The trained model was integrated into the iOS app. It provided inline "ghost text" suggestions triggered by the spacebar.
  • Interaction Design: The system allowed for partial acceptance of suggestions (accepting 1 word vs. a full paragraph), giving the user granular control.
  • Data Collection: The author used the app as his primary AAC device, logging usage analytics (acceptance rates, word counts) and completing structured diary entries every few days.

3. Key Contributions

1. Empirical Insight into Ultra-Personalization: The study provides a grounded, longitudinal account of how an LLM trained on a single user's entire communication history reshapes their agency, identity, and privacy boundaries.
2. Identification of Tensions: It surfaces critical trade-offs in personalized AI:
   • Agency: Personalization can constrain agency through self-censorship during data collection and "steering" during usage.
   • Identity: While models can amplify cultural identity (e.g., code-switching), they risk erasing authentic "messy" aspects of speech (humor, slang) if filtered too strictly.
   • Privacy: Models can breach Contextual Integrity by resurfacing intimate or context-specific details (e.g., religious background, family history) in inappropriate social settings.
3. Positionality in Research: The paper highlights the value of the "user-researcher" perspective in revealing subtle, lived experiences (like the feeling of being "watched" by a logging system) that external observation might miss.

4. Key Results & Findings

• Agency and Self-Censorship:

  • Data Collection: The knowledge that data was being logged caused the author to self-censor, altering his natural communication style to create a "cleaner" dataset.
  • Usage: In structured contexts (academic meetings), suggestions enhanced fluency and agency. However, in rapid or niche contexts (debugging code, casual bar talk), the model often hallucinated or steered the conversation in unwanted directions, reducing agency.

• Identity and Cultural Nuance:

  • The model successfully captured the author's multicultural identity, including Spanglish and Argentine slang (e.g., "che boludo").
  • In some instances, the model acted as a co-constructor of identity, completing sentences with culturally specific terms the author would have used, reinforcing his authentic voice.
  • Conversely, the filtering process in Phase 2 resulted in a model that lacked the "dark humor" or "messiness" of the author's true personality.

• Privacy and Contextual Integrity:

  • Contextual Breaches: The model occasionally resurfaced private biographical details or religious markers in professional or casual settings where they were inappropriate, violating the "contextual integrity" of the conversation.
  • Forced Intimacy: Because the author uses a "text-showing" style (showing the screen to interlocutors while typing), the visibility of AI suggestions led to "forced intimacy." Interlocutors sometimes saw private details or AI-generated text before the author committed to it, leading to confusion about who was speaking (the "AI Ghostwriter Effect").

• Usage Analytics:

  • The author accepted suggestions in 153 out of 6,564 messages.
  • When accepted, the average length was 7.5 words, with some instances of accepting up to 55 words.
  • Acceptance rates were high (100% of words accepted in 114 cases) when the context aligned, but the model was frequently rejected in rapid-fire or technical conversations.

5. Significance and Design Implications

The paper argues that ultra-personalized AAC cannot be a static, one-time training event. Instead, it proposes a shift toward adaptive, user-steered ecosystems.

• Context-Aware Surfacing: Systems must determine when to surface recommendations based on social norms and context (e.g., hiding suggestions in professional settings if they risk privacy breaches).
• Modular Identity: Rather than a single monolithic model, future systems should use modular, context-aware models that can switch between different facets of a user's identity (professional vs. casual vs. intimate).
• User Tuning: Users need granular control to tune the model's behavior, including:
  • Tuning Suggestions: Allowing partial acceptance and "anchors" to guide the AI.
  • Tuning Intent: Detecting writing styles (planned vs. exploratory) to adjust verbosity.
  • Tuning the Model: Providing no-code interfaces for users to adjust how the system learns and behaves over time.

Conclusion:
Ultra-personalized AI holds immense potential to restore fluency and expressivity for AAC users, but it introduces complex ethical risks regarding privacy, authorship, and the authenticity of the user's voice. The study concludes that successful deployment requires moving beyond technical optimization to design systems that continuously negotiate the boundaries of agency, identity, and context with the user.