Access That Lasts

World Usability Day 2026 · Designing a Sustainable Future with Ethical AI

Mick McQuaid

University of Texas at Austin

13 Nov 2026

Two years ago

In 2024 I stood here and said “AI isn’t helping (much)”

  • I would be half right to say it again
  • Vendors still sell magic fairy dust for compliance
  • But disabled people adopted general-purpose AI faster than almost anyone
  • Today’s question is different: not does it help but will it last

The thesis

  • AI is the biggest accessibility advance in decades
  • It is also the fastest way to replace real access with an unverified, impermanent approximation
  • A future is not sustainable if it quietly removes access for the people who need it most
  • It is not ethical if it does so while claiming inclusion
  • UX is the one place where a disabled person’s actual experience becomes visible to the organization

The promise

What AI is doing for disabled people right now

  • On-demand image and scene description
  • Plain-language rewriting and summarization
  • Voice-first interaction and live captions
  • Document and code remediation at scale—but with an important caveat, Cipriano et al. (2026)
  • Assistive use is one of the strongest justifications for the energy and water cost of large models. It is also often the first feature cut.

The problems

Bias and misrecognition

  • Atypical speech, gait, faces, and bodies are underrepresented in training data
  • Speech recognition for dysarthric speech lags far behind typical speech
  • Vision models misread wheelchairs, prosthetics, and service animals

Automation replacing human access

Vendor advertisement: Mitigate your risk of digital accessibility lawsuits and ensure your content is usable by everyone by making your PDFs compliant with WCAG, Section 508 and the ADA. Buttons read Equidox AI and Equidox Software.

  • Auto-captions, alt text, and signing avatars sold as “good enough”
  • AI used to justify cutting captioners, interpreters, and support staff
  • Captioners and interpreters are infrastructure: slow to rebuild once gone, and the automation’s quality bar depends on them

The verification problem

  • A blind user receives a fluent, confident, wrong image description
  • A plain-language summary drops a legally important sentence
  • Trust is a depletable resource. Each confident error burns some trust.
  • Autonomy, too: AAC prediction nudges a user’s voice toward the model’s norms, and benefits algorithms cut care hours with no explanation and no appeal (Eubanks 2018)

Unsustainable access

A long white arched corridor reached only by a flight of steps, with no ramp

  • Capable assistive AI is increasingly subscription-based and cloud-dependent
  • Fixed incomes and low bandwidth lock people out
  • Model deprecations and silent behavior changes break assistive workflows without warning
  • Access that depends on a vendor’s roadmap is not access
  • Accessibility debt: features ship faster than anyone can retrofit, and a ramp added later costs more than a ramp designed in

The four tests

Every AI feature must pass all four

  • Accessible. Can they operate it with their assistive technology?
  • Checkable. Can they tell when the output is wrong?
  • Human. Can they reach a person as easily as they reached the AI?
  • Durable. Will it still work in five years, offline, without a new subscription?
  • Fail one and it is not accessible, whatever the compliance report says

What you can do

Research

  • Recruit disabled participants by default, across disability types, and pay for access needs
  • Hire disabled researchers and designers. Participation at the decision level, not just the study level (Bennett and Rosner 2019)
  • Watch how people decide whether to trust AI output
  • Ask what human work the feature replaces

Design

  • Show confidence, provenance, and what the model could not see
  • Keep a human reachable, and make escalation as findable as the AI path
  • Keep suggestions visibly optional and distinct from the user’s own words
  • Test the AI interface itself with screen readers, switch access, voice control, and magnification: streaming text, chat, and voice-first are frequently inaccessible
  • Prefer small on-device models where they meet the need: greener, cheaper, and they work offline

Organization

  • Disaggregate quality metrics for disabled users and track regressions across model updates
  • Refuse the “good enough” argument when it is used to cut human services
  • Write disability bias and durability into procurement: disaggregated performance data, deprecation notice periods, offline fallbacks, data export, exit plans
  • Get disability named in your AI ethics principles and in your sustainability commitments

Two collaborations,
twenty years apart

Then: HathiTrust and blind readers

  • HathiTrust made out-of-copyright books available to everyone
  • They also made in-copyright works available to blind people
  • Publishers sued, claiming that the HathiTrust technology would be misused to give sighted people access
  • Should you let illegal misuse stand in the way of accessibility?
  • The courts sided with HathiTrust
  • History has supported their decision

Now: Meta and the National Federation of the Blind

  • A holy grail for blind people would be the ability to identify the people they know when they walk into a room
  • The NFB worked with Meta to realize that dream through Meta glasses
  • A privacy-focused NY Times columnist got wind of the project and opined that the technology would be misused by sighted people to invade privacy
  • Meta quickly pulled the plug on the project, lacking the balls that HathiTrust had shown almost twenty years ago
  • Should you let illegal misuse stand in the way of accessibility?

What the parallel teaches

  • According to accessibility pioneer Greg Vanderheiden, we’re slowly losing the war for digital accessibility
  • He says it’s because new systems are coming online faster than we can design in accessibility
  • But is it also a question of diminishing will to resist obstacles?
  • Hopefully, the midterms are a positive sign!

Close

Everyone who lives long enough

Inclusive Design chart of permanent, temporary, and situational disability across touch, see, hear, speak, and cognitive: one arm, arm injury, new parent; blind, cataract, distracted driver; deaf, ear infection, bartender; non-verbal, laryngitis, heavy accent, adhd, fatigue / depression, open office landscapes
  • Designing for disabled users today is designing for the future population
  • That is the literal definition of sustainable design

One thing this month

  • Pick one item from the three “what you can do” slides and do it before December
  • Sustainable, ethical AI is measured by whether it still works for a disabled person in five years, not by the demo
  • Accessible. Checkable. Human. Durable.

Please contact me!

email mcq@utexas.edu to share further ideas and questions

visit https://mickmcquaid.com and click on Accessibility

Thanks for listening!

END

References

Bennett, Cynthia L., Cole Gleason, Morgan Klaus Scheuerman, Jeffrey P. Bigham, Anhong Guo, and Alexandra To. 2021. “‘It’s Complicated’: Negotiating Accessibility and (Mis)representation in Image Descriptions of Race, Gender, and Disability.” In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery. https://doi.org/10.1145/3411764.3445498.
Bennett, Cynthia L., and Daniela K. Rosner. 2019. “The Promise of Empathy: Design, Disability, and Knowing the "Other".” In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery. https://doi.org/10.1145/3290605.3300528.
Carik, Buse, Kaike Ping, Xiaohan Ding, and Eugenia H. Rho. 2025. “Exploring Large Language Models Through a Neurodivergent Lens: Use, Challenges, Community-Driven Workarounds, and Concerns.” Proceedings of the 2025 ACM International Conference on Supporting Group Work (GROUP ’25). https://doi.org/10.48550/arXiv.2410.06336.
Cipriano, Colleen F., Yichun Zhao, Miguel A. Nacenta, Kotaro Hara, and Jaylee Soh. 2026. “Q&a or Document-Based? The Effects of Interface Type on How Screen Reader Users Access Interconnected Documents: Q&a or Document-Based for Screen Reader Users?” Proceedings of the 28th International ACM SIGACCESS Conference on Computers and Accessibility (New York, NY, USA), ASSETS ’26. https://doi.org/10.1145/3797867.3829036.
Eubanks, Virginia. 2018. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin’s Press.
Heidt, Amanda. 2024. “‘Without These Tools, i’d Be Lost’: How Generative AI Aids in Accessibility.” Nature (London) (London) 628 (8007): 462–63.
Overlay Fact Sheet signatories. 2021. Overlay Fact Sheet: An Open Letter from Accessibility Practitioners. https://overlayfactsheet.com.
Trewin, Shari, Sara Basson, Michael Muller, et al. 2019. “Considerations for AI Fairness for People with Disabilities.” AI Matters 5 (3): 40–63. https://doi.org/10.1145/3362077.3362086.
U.S. Equal Employment Opportunity Commission. 2022. The Americans with Disabilities Act and the Use of Software, Algorithms, and Artificial Intelligence to Assess Job Applicants and Employees. https://www.eeoc.gov/laws/guidance/americans-disabilities-act-and-use-software-algorithms-and-artificial-intelligence.
Whittaker, Meredith, Meryl Alper, Cynthia L. Bennett, et al. 2019. Disability, Bias, and AI. AI Now Institute. https://ainowinstitute.org/publication/disabilitybiasai-2019.

Colophon

This slideshow was produced using quarto

Fonts are Fira Sans, Fira Sans Light, and Victor Mono Nerd Font

Math is set in Fira Math via MathJax 4