LaMPost: Design and Evaluation of an AI-assisted Email Writing Prototype for Adults with Dyslexia
Steven M. Goodman, Erin Buehler, Patrick Clary, Andy Coenen, Aaron Donsbach, Tiffanie N. Horne, Michal Lahav, Robert MacDonald, Rain Breaw Michaels, Ajit Narayanan, Mahima Pushkarna, Joel Riley, Alex Santana, Lei Shi, Rachel Sweeney, Phil Weaver, Ann Yuan, Meredith Ringel Morris · 2022 · Proceedings of the 24th International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS '22) · doi:10.1145/3517428.3544819
Summary
This paper introduces LaMPost, a prototype email-writing interface powered by Google's LaMDA large language model, designed to support adults with dyslexia through AI-assisted writing features. The research team, which included members with lived experience of dyslexia, engaged in over a year of participatory research with the dyslexia community before building the system. LaMPost augments a standard email editor with three LLM-powered features: "Identify Main Ideas" generates a visual outline of each paragraph's main idea with the option to auto-generate a subject line; "Suggest Possible Changes" offers AI suggestions for improving a selected passage; and "Rewrite My Selection" allows users to select text and provide custom instructions for how the AI should rewrite it (e.g., "to be more formal," "to be shorter"). The system incorporates accessibility considerations including an 18pt sans-serif font, 140% line spacing, icon-paired buttons, text highlighting near the cursor, and a "read aloud" feature using the Web Speech API. The team also experimented with several other LLM-based features during development, including a chatbot-style conversational drafting interface and automatic outline-to-email expansion, but found these produced unreliable results due to hallucinations and context window limitations. The evaluation involved 19 adults with dyslexia in a 75-minute remote session that included background interviews, a hands-on writing exercise, and post-use rating scales.
Key findings
The "Rewrite My Selection" feature was rated most useful overall (avg. 5.26/7), with nine participants selecting it as the most useful feature. Participants valued its ability to help them find satisfying wording that matched their intended tone and style, with custom instructions like "make it shorter" or "more business-like." The subject line generation received surprisingly strong positive response — many participants reported always leaving subject lines blank because nothing felt right. The "Identify Main Ideas" feature was valued primarily as a validation tool, confirming that the email communicated what the writer intended. However, significant concerns emerged around LLM accuracy: hallucinations appeared in both the Rewrite and Suggest features, with the model sometimes adding fabricated details to rewritten passages. Participants experienced a "paradox of choice" when presented with too many rewrite options (up to 15), finding it cognitively demanding to parse and compare similar options — they recommended reducing to 3-5 choices. A key surprising finding was that the between-subjects AI-framing manipulation (presenting the system as "AI-powered" vs. an "enhanced" editor) had no significant effect on any rating including usefulness, satisfaction, self-expression, self-efficacy, autonomy, or control. Seven participants raised privacy concerns about the system reading their personal email data. Despite enthusiasm for the features, participants concluded that 2022-era LLMs did not yet meet the accuracy and quality thresholds needed for reliable daily use by writers with dyslexia.
Relevance
This paper is a landmark early exploration of LLMs for accessibility, published at a time when generative AI tools were just beginning to reach mainstream adoption. Its findings remain highly relevant as AI writing assistants have since become ubiquitous. The identification of specific writing challenges faced by adults with dyslexia — organizing ideas, matching tone, word retrieval, and the "fear of the blank page" — provides a requirements framework for any AI writing tool targeting this population. The paradox of choice finding is particularly important: presenting many AI-generated options, a common design pattern, creates additional cognitive load for users who already struggle with reading and parsing text. The null result on AI framing is noteworthy — it suggests that for this population, the practical utility of writing support matters more than whether users know AI is involved. For organizations deploying AI writing tools, the paper highlights that hallucinations and inaccuracy are not merely quality issues but trust-breaking events that are especially harmful for users who may have difficulty independently verifying AI-generated text. The privacy concerns raised foreshadow ongoing debates about AI systems processing personal communications.
Tags: dyslexia · large language models · AI writing support · email · writing challenges · human-AI interaction · neurodiversity · text generation · accessibility tools