When Disability Disclosure Travels: Memory, Privacy, and Contextual Integrity in Conversational AI
In short
AI assistants now remember what we tell them. I interviewed 12 disabled adults about what they share with assistants such as ChatGPT, Claude and Gemini, and what happens to it afterwards. Memory spared people from explaining themselves again, but it also carried disability into conversations where it did not belong.
Why this work
Telling someone about a disability has always been a trade between getting access and risking stigma, decided one audience at a time. An AI assistant feels like a patient helper that does not judge. It is also a company that stores what is said, and its memory carries a disclosure into later conversations.
We wanted to know what a remembering assistant should know, across many kinds of disability.
What we did
I interviewed 12 adults in the United States with self-identified disabilities. They included people who are blind or have low vision, people with mobility or physical disabilities, people who are d/Deaf or hard of hearing, and people with chronic illness, pain or fatigue. Each used an AI assistant at least twice a week, or had cut back.
Interviews ran about an hour on Zoom. Two of us coded every transcript. We read the themes through contextual integrity, the idea that information carries the norms of the context it came from.
What we found
People disclosed by need, not by name. Half first described an instruction, such as how they want information presented, not a diagnosis.
One chat has two recipients. Half found it easier to tell an AI than to tell people. Three kept the company behind the assistant in mind, and shared less with it than they would with a trusted person.
Memory brought relief and drift. Most welcomed being remembered. But remembered details were over-applied, pinned on the wrong person, or leaked from one topic into another.
Assistants did not always act on what they knew. One participant who cannot stand was cautioned about standing up too fast. Another, who has muscular dystrophy, was told to take a walk.
People did their own boundary work, such as keeping separate accounts for separate parts of life, or asking an assistant what it knows and correcting it. This works, but it takes effort and technical skill.
What it means for design
Honor the stated need, and do not ask about disability. Check recommendations against the limits a person has stated. Show what is stored and where it came from, tie it to its original context, and say when a stored fact is shaping an answer.
A remember-this switch on every message was not the answer: half of those who discussed it turned it down. And because memory is itself an accessibility feature, weakening it for the sake of privacy would fall hardest on disabled users.
Cite this paper
Atieh Taheri, Mahya Tazike, Patrick Carrington, and Jeffrey P. Bigham. 2026. When Disability Disclosure Travels: Memory, Privacy, and Contextual Integrity in Conversational AI. arXiv:2609.22720. https://arxiv.org/abs/2609.22720
BibTeX
@misc{taheri2026disabilitydisclosuretravelsmemory,
title = {When Disability Disclosure Travels: Memory, Privacy, and Contextual
Integrity in Conversational AI},
author = {Atieh Taheri and Mahya Tazike and Patrick Carrington and Jeffrey P.
Bigham},
year = {2026},
eprint = {2609.22720},
archivePrefix = {arXiv},
primaryClass = {cs.HC},
url = {https://arxiv.org/abs/2609.22720}
}