Your UI is not my UI

Simon Lenz & Matthew Seccafien

AI & Accessibility today

Hypothesis

Chatbots and agents are a total revolution for accessibility

User research

User Interviews

We interviewed six people with disabilities

Melody

AI as an Accessibility “Assistant,” not a Replacement

“I do use AI, but only as an assistant for what I'm looking for. Not to do the searching from scratch.”

A blurred figure of a woman walking down a stairwell, reading her phone

Matthew

Accessibility AI should be like training wheels, not autopilot

“[…] you have to put your name to things. If it went ahead, did a bunch of tasks, emailed people, and I couldn't really take it back — that feels … uncomfortable, giving it that much autonomy.”

Sinead

AI can explain the cause, not just adapt the interface

“I was really struggling one day … but then, interestingly, it came up with a really insightful reason: it was the drop in barometric pressure. The change had been so fast. And it said that's really common with people with brain injuries.”

Sinead had a brain injury five years ago — came to disability later in life. Screen time is hard; reading a lot in one day is hard.

Her setup is a stack of fighting extensions: a plugin that bolds half of each word breaks the one that overlays numbers for voice control; the number-grid alternative is invisible under her dark-mode plugin. It doesn't work in Safari, and it collapses the moment she also wants to watch a lecture.

She has strong preferences (she likes Gwyneth Paltrow's voice in her screen reader) — but can't get these tools to work together.

AI providing information in an interface that is highly compatible and tailored for the user — not just adapting a layout, but making sense of what's happening to them.

No Revolution

Agents & chatbots are very helpful tools
Simple interface is very accessible
Trust & Reliability is a large problem
We want assistants and not replacements

Allana

AI should only remove the tax that a disability imposes on everyday tasks

“Part of the enjoyable thing in life is struggling through things. Having everything taken care of for you takes away a bit of the fun.”

Rose

Everyone's settings are different and incompatible

“I live in a disabled house […] all our settings are different, even across our iPads. It's actually an in-house joke: I can't look at my husband's iPad, he can't look at mine, and the kids hate using either.”

Myron

AI can provide an interface where you are already a power user

“If I had an AI box, I could tell it to line up all my headers from cell A1 to A20 and just send it. That would save so much time. That would be a life changer for me.”

The limits of adaptability

Today's model

Opportunities

Setups are very personalized
Adaptations conflict with each other
Configuration costs energy
Standardization has limits

→ One interface can't fit the variety of needs of its diverse users

Personalization over adaptability

Your UI is not my UI

Personalization over adaptability

Your UI is not my UI

Create specific UIs on the fly for your needs

Adapt to your current task
Adapt to your preferences
Adapt to your accessibility needs
Adapt to your circumstances
Adapt to your state of mind
Adapt to your external environment

A UI should be the result of two things:

what you're trying to do,
and who you are.

An app for the exact same task needs to look different depending on the condition, state, environment and preferences of the user. In one word, the interface needs to be empathetic to the person in front of it.

The empathetic UI meets you where you are, so the adaptation work disappears. [While we talk here about this in an accessibility context, we think this is a generally desired property that goes beyond the topic.]

Where this goes: launching an app is creating an app, just for you, in that moment. Generated on the fly, instead of adapted with layered tools.

Demo: same task, two people

Person A

Stressed by long paragraphs, prefers dark backgrounds, standing on a subway with headphones on, one hand free, headache from a barometric pressure change.

[ placeholder — calm, minimal, dark-mode UI ]

Person B

Has a tremor, avoids precise pointer input, prefers keyboard interactions, at home in a relaxed setting writing an email to friends.

[ placeholder — large keyboard-first targets ]

Empathy is our point

[UI] = f([Intent], [Empathy])

Observations of existing state. The industry is already doing this

[here screenshots of those examples might fit]

Google Search AI Mode — custom layouts assembled per query, not served from a template
Artifacts — every major AI tool now generates working interfaces inline, on demand

What is gained?

Burden of adaptation is shifted from the user to the application
Users not at the mercy of hopefully compatible application following standards

Launching is creating

What if launching an app means creating an app?

Focus shifts from UI to API

Launching an app is programming an app
Focus on content and functionality (APIs), not UI
The interface becomes disposable — generated fresh, every time

Applications are no longer a commodity

A hand-crafted UI stops being the default way software is distributed
Disabled users are just the ones suffering most from this today
Personalized UIs would help everyone, not only accessibility use cases

Why have we always had one-size-fits-all interfaces? It's not a design failure — it's economics. Building a custom app for every individual need was never affordable, so we abstract what a big population has in common and ship one reusable solution. That's why there's no dinner-party app, no family-travel app, no grant-writing app — everyone does those differently. Instead we get rigid apps (the developer decides what matters — Airbnb won't show “pet-friendly” on the map) or bloated ones (everything dumped in). AI changes that economics: for the first time, the made-for-one interface is affordable. That's the “why now.”

And it's not a new idea — the field's founders chased it for decades. Engelbart, in the 1960s, complained that we design for a “statistical population” with no attempt to fit the individual. David Karger later put it: “no developer can predict all the ways a user will want to manipulate information.” They were right, and stuck — the idea was lovely but impractical, because ordinary people had no way to build it. That's the part that just changed.

Google Research frames the shift well: LLMs already turned a finite library of texts into an infinite one — generated on the spot for any need. Interfaces are next: instead of picking from a finite library of apps, you get the right interface generated on the spot, for you.

For accessibility this lands hardest. No standards body can anticipate Sinead's exact combination of a brain injury, dark-mode plugin, and voice-control grid. The person who understands that interface best is Sinead — and for the first time, it can be made for her, in her context.

Revisiting: 8 months later

Rose tried to build a bespoke Steam library manager

It didn't really work out

Lack of support
Lack of knowledge
Still too time consuming
Too much tinkering required

We replaced configuration toil with developer toil.

Is Rose's failure the idea, or the moment?

We think it's the moment.

Whose context is it?

A UI generated for you needs to know a lot about you
For our interviewees that's the most sensitive data there is — health, disability, cognitive state
Generated interfaces only work if that context can stay private
On-device / local AI is the path that keeps it that way

“But a UI that keeps changing breaks screen readers.”

Personal is not the same as volatile

Screen readers and muscle memory depend on stability
Personalize the content and what's pre-selected — never relocate the controls
Generated once for you — then remembered and reused
Stability is empathy for anyone who navigates by memory

The sharpest objection, met head-on: if interfaces are generated fresh every time, don't you destroy the thing a screen-reader user relies on — consistency, muscle memory, the button being where it was yesterday? A constantly re-rolling UI would be an accessibility disaster.

But personalization isn't volatility. The model is memory: generated once, for you, then remembered and reused. And that's itself empathy — an interface that understands a screen-reader user knows they navigate by memory, so it holds still. Meanwhile today's one-size-fits-all UI changes underneath you whenever a vendor redesigns it — that's the Ravelry change that put Rose in bed for three days.

The rule: personalize the content, not the structure. Don't move the button — change what's shown and pre-selected. Not “different every time,” but “different for each person, then dependable for that person.”

Closing

Key takeaways

Today we adapt one majority interface — it can't serve the variety of humans
An interface should account for both the task and the human doing it
With generated UIs, that's economically viable for the first time in history

Back to Sinead

No fighting plugins. No dark grid she can't see. No setup that collapses when she opens a lecture. Just an interface that already understands her — and already fits.

Let's end where we started — with a person, not a paradigm. Remember Sinead: the bolding plugin fighting the voice-control grid, the grid lines invisible under dark mode, the whole fragile stack collapsing the moment she wants to watch a lecture. Every layer she added to adapt a hostile interface was one more thing that could break.

Now imagine the interface simply arrived understanding her — her contrast, her reading load, her voice control — the result of what she's trying to do and who she is. Not adapted. Not fought with. Empathetic to her from the first moment. That's not a nice-to-have for Sinead. That's the difference between software she can use and software that puts her in bed for three days.

That's the future we want to build. Questions?

Thank you!

Q&A

Backup — Q&A

Anticipated objections

“Doesn't this just hand more power to whoever owns the AI?”

A generated UI can be empathetic to you — or to whoever owns it.

“If everyone's UI is different, don't we lose common ground?”

Real cost. Keep the periphery visible — personal shouldn't mean sealed.

“Will people actually use this, or just accept the default?”

The empathy framing sidesteps it — the default already fits.