Our paper is out at ACM ASSETS 2026: Reconfiguration of Ability

| Katsuki Ono

ASSETSpaperaccessibilityLLMProject VOICE

My single-author paper has been published at ACM ASSETS 2026, the 28th International ACM SIGACCESS Conference on Computers and Accessibility (October 25–28, 2026, Vila Nova de Gaia, Portugal). It is open access under CC BY 4.0, so anyone can read the full text.

Reconfiguration of Ability: An Autoethnography of LLM-Mediated Development by a Developer with Severe Physical Disability
https://doi.org/10.1145/3797867.3828990

From a bed, through a single switch

I cannot speak, type, or move my limbs. Only my head moves. A switch mounted beside it is my only interface with a tablet: a cursor scans across a grid of characters, and I press the switch when it reaches the one I want. Typing a sentence like "Please report on the current progress of model accuracy" takes about 1 minute and 28 seconds.

Even so, over the past year I built eleven versions of a text-prediction model that writes in my own voice, plus an auxiliary model to synthesize its training data. I did not write a single line of code by hand. I told AI what I wanted in natural language, read the code and results it produced, judged them, and gave the next instruction — what is now called "vibe coding." The model was developed as part of Google's Project VOICE.

The paper is an autoethnography — a method that uses the researcher's own experience as primary data — of that development, drawing on development logs from October 2025 to March 2026 and on reflective narratives.

What eleven versions taught me

  • v1–v2: Adding a fixed instruction ("predict the continuation as a keyboard prediction") to every training sample stabilized the output immediately. How the task is framed mattered more than data volume.

  • v3–v4: An attempt to generate multiple candidates failed. Metrics said the predictions were correct, but five near-identical candidates are useless. I could call it a failure only because I am the person who uses that keyboard every day.

  • v5: The predictions came out consistently casual. The data came only from personal accounts — business accounts were excluded for confidentiality — and was dominated by messages to family and close friends. The model was a mirror showing only part of my face.

  • The auxiliary model: A generic LLM writes business text that is correct but sounds like anyone. So I fine-tuned a separate model on my own writing and had it synthesize 16,000 business-register samples in my voice. I built a tool to build the tool.

  • v6–v11: Fewer epochs, broader data. I learned that "I" am not a single register but a repertoire.

In a blind comparison of the final model (v11), all four evaluators — two family members and two acquaintances — identified its outputs as sounding like me. What stayed with me most, though, was my own reaction: reading its predictions felt like rereading messages I had written and forgotten, an embarrassment at being seen too clearly.

Reconfiguration of ability

The paper names this experience reconfiguration of ability: neither the "restoration" of function nor the "removal" of barriers. It has three features:

  1. Persistence of impairment: Across all eleven versions my body did not change. Same bed, same switch.

  2. Non-removal of barriers: Conventional development environments are still inaccessible to me. I simply no longer need them.

  3. Redefinition of the ability itself: What development demands shifted from typing speed and syntax recall to articulating intent, evaluating output, and setting direction.

This differs from compensation (doing the same thing by other means) and from accommodation (adapting the environment), because the activity itself changes. Good engineers have always been valued for design and judgment — but turning that judgment into working software used to require passing through a gate of physical code manipulation. LLMs removed the gate.

I did not become able to type. I became able to design.

For people who build assistive technology

  • A single conversational entry point: Rather than making each tool — editor, terminal, email — accessible one by one, a single natural-language entry point can reach all of them.

  • Identity-preserving personalization: Not word lists or stored phrases, but a person's phrasing, sentence rhythm, and formality. Users should control what the model learns and what counts as "sounding like me."

  • Two axes: Adapting existing interfaces to diverse bodies, and removing the need for those interfaces altogether. Neither alone reaches everyone.

Limitations

This is a single case. I had self-taught experience in LLM post-training and fluent written Japanese; for people whose disabilities affect language, the same path does not apply. All eleven versions were built on a commercial cloud model whose pricing and deprecation are outside my control — which is why I am also pursuing small on-device models (Gemma 3 270M).

People sometimes ask, "Did you really build that?" The body that typed the characters was an AI's. But I envisioned its purpose, designed it, diagnosed its failures, and directed every revision. This paper was written the same way.

Links

‹ Back to Blog