U2xAI
…
ConceptIntermediateDesigning for Uncertainty & Trust / UX for uncertainty and confidence display / #20

What is the accessibility consideration for confidence indicators?

GUARD name who can't perceive the cue, not just who built it

Meadowlark Public Library runs Stackwise, an AI reference assistant on its kiosks and website. It answers a patron's question and marks each answer with a confidence cue, so patrons know whether to trust it or ask a librarian. Selin Karaca manages Digital Services there. Junko Adeyemi has used the library for eleven years and reads every screen with a screen reader.

The direct answer
Never encode confidence in color alone. Pair every color cue with a plain text label, an icon shape, and a spoken-equivalent line a screen reader actually announces, so a patron who can't see the dot still gets the same warning a sighted patron gets.
Do this, in order
  1. Give every confidence cue a text label and a spoken-equivalent line, never color alone.Why: color carries zero meaning to a screen reader, no matter how visible it is on screen.
  2. Test the actual screen reader output on a real device, not just the visual design.Why: a badge can pass a contrast-ratio check and still say nothing new out loud.
  3. Track completion and follow-up rates split by assistive-tech usage.Why: the patrons losing this signal will not file a complaint about something they were never told existed.
  4. Name both people harmed by a design gap, not just the team that built it.Why: a design review that only asks "does this look clean" never asks who it looks clean to.
  5. Treat any removed accessible label as a regression, never a simplification.Why: "cleaner" for one group can quietly mean "silent" for another.
  6. Leave the sighted visual design's colors and icons alone.Why: the fix is adding a missing channel, not removing the ones that already work.

How to answer this, stage by stage

Nobody is grading whether you can name a WCAG rule number. They're grading whether you can find the patron your design quietly stopped talking to.

Stage 1
Scope it to one product, one pair of people
Say it like this
"I'll answer this for Stackwise, Meadowlark Public Library's reference assistant, and its confidence badge, the one every patron sees or is supposed to hear."
Why this works
Keeps "the accessibility consideration" from staying a vague compliance term with no product behind it.
Stage 2
Say your structure out loud
Say it like this
"I'll use GUARD. Groups, who's affected. Unequal, where the harm actually lands. Ability to contest, who can't push back. Reduce, the real design change. Detect, how you'd catch it in production."
Why this works
Shows a real method for a fairness question, not a single good instinct.
Stage 3
Name both groups, not just the one you can see
Say it like this
"There's the team that built the badge, and there's the patron using a screen reader who never gets what the badge is showing. Most reviews only ever look at the first group."
Why this works
Puts a real person, not an abstract "accessibility users" category, into the answer.
Stage 4
Say where the harm lands unevenly
Say it like this
"A sighted patron loses nothing here. A patron using a screen reader loses the entire warning, every single time, on every answer Stackwise is actually unsure about."
Why this works
Turns "accessibility matters" into a specific, unequal cost instead of a general good intention.
Stage 5
Ask who never gets to push back
Say it like this
"Junko never got a support ticket to file, because she never knew the label was supposed to be there in the first place. You can't flag the absence of something you were never told existed."
Why this works
This is GUARD's hardest step, and the one that separates a real answer from "we care about accessibility."
Stage 6
Give the concrete design change
Say it like this
"Never let color alone carry the meaning. Pair it with a text label and a spoken line the screen reader actually reads out loud, every time."
Why this works
This is the direct answer, said as a specific build decision instead of a policy statement.
Stage 7
Say how you'd detect it in production
Say it like this
"Track follow-up rate on low-confidence answers, split by assistive-tech flag. If screen-reader patrons never ask a follow-up question, that's not calm trust, that's a missing signal."
Why this works
Shows you're not relying on a complaint that, by design, was never going to arrive.
Stage 8
Close on the one line
Say it like this
"An accessible confidence cue isn't one look everyone can see. It's the same warning, reaching every patron through whatever channel actually works for them."
Why this works
Restates the direct answer in one breath, ready for a follow-up.

Let's learn

Stackwise is an AI assistant Meadowlark Public Library runs on its kiosks and website. A patron asks a question, Stackwise answers, and marks the answer with a confidence cue, so patrons know whether to trust it or ask a librarian.

Before Stackwise, a patron asked a librarian directly, always got a person's real judgment, and sometimes waited five or ten minutes during a busy afternoon. Now Stackwise answers about 900 questions a day instantly, freeing librarians for harder requests. About 27 of those sessions a day, roughly 3 in 100, come from patrons using a screen reader.

Hand sketched metaphor scene titled Two people, one lever. Left, a purple person icon labeled Operator, caption can flag can override. Right, an amber person icon labeled Subject, caption gets the answer no lever.
Selin's team can change the badge any time. A patron using a screen reader can only receive whatever the badge chose to say.

Here's the turn: a visual redesign meant to make the kiosk screen "cleaner" for sighted patrons replaced a spoken confidence label with a plain color dot, and quietly changed its screen-reader description to a generic "status indicator," with no actual value inside it. Sighted patrons could still read green, amber, or red at a glance. Patrons using a screen reader got nothing new at all, and had no way to know anything was even missing.

Patrons who asked a librarian follow-up after a low-confidence Stackwise answer
60% 30 0 57% 57% Sighted patrons 5% 54% Screen reader patrons
Sighted patrons never noticed a gap. Screen reader patrons were missing the entire warning, before the spoken label came back.

At its worst, a patron acts on a Stackwise answer that Stackwise itself was unsure about, with nothing telling her to double check, simply because the one channel that would have warned her never reaches a screen reader at all.

Hand sketched flow diagram titled Where the appeal should be, and isn't. Five boxes: Stackwise answers, Confidence shown sighted only, Nothing for screen reader users highlighted, No flag ever raised, Answer stands unchecked.
The gap sits in the third box. Everything after it follows from that one missing step.
The decision I would take back Selin's team decided, during the redesign, to drop the specific spoken label in favor of a cleaner-looking badge, because the old label read as clutter in an accessibility check that was actually only testing the visual badge's color contrast, not what a screen reader announced. That check passed. It just wasn't checking the thing that mattered.

What I would leave alone: the sighted visual design itself, the actual colors and icon shapes chosen, doesn't need a rebuild. The gap was never what sighted patrons saw. It was what got left out for patrons who couldn't see it at all.

The lesson: an accessible design isn't one look everyone can see. It's the same warning, reaching every patron through whatever channel actually works for them.

Now here is the same thing as a story

The short version above is what you'd say defending this fix to Meadowlark's library board. Read this one for how quietly the gap actually opened.

The reference desk at Meadowlark Public Library used to have a line most Thursday afternoons, patrons waiting to ask a librarian something a catalog search couldn't answer.

Junko Adeyemi has used the library for eleven years, and reads every screen, at home and on the library's own kiosks, with a screen reader. She was one of the first patrons to try Stackwise on the day it launched, and for months it worked exactly as well for her as it did for anyone else, the confidence label read aloud right alongside the answer.

Selin Karaca manages Digital Services at Meadowlark and led a redesign meant to make Stackwise's kiosk screen feel calmer, less like a form, more like a person talking. Patrons loved the cleaner look. The badge shrank from a labeled box to a single color dot with an icon.

Knowledge spark: what does a screen reader actually announce? A screen reader reads out the text and labels attached to what's on screen, not the picture itself. A colored dot, on its own, has nothing attached for it to read. If a designer doesn't add a specific label describing what that dot means, the screen reader either stays silent or reads something generic, like "button" or "status indicator," which tells a patron nothing new.

Nobody on the design team tested the new badge with a screen reader before shipping it, since the visual version had already passed a contrast-ratio accessibility check. That check answered a real question. It just wasn't the question that mattered here.

Hand sketched labeled parts diagram titled What's in an accessible confidence cue. A document icon at the center labeled Confidence Cue, with four callouts around it: color, icon shape, text label, screen reader text.
Four channels. The redesign kept two of them and quietly dropped the two that mattered most to Junko.

Six months later, Meadowlark's state library association ran its annual accessibility review and pulled ten random kiosk sessions to test with a screen reader. Nine were fine. The tenth was Stackwise's confidence badge, and it read only "status indicator, button," nothing else, on every single answer, sure or not.

Junko had used Stackwise regularly to check large-print book availability for her book club, and had quietly stopped noticing anything different, since nothing in what she heard ever changed between a confident answer and an unsure one. She had no way to know the cue she used to rely on had gone silent months earlier.

We did not take a color away from Junko. We took away the only warning she ever had.

The auditor's finding wasn't really about Junko's book club question, mostly harmless as it was. It was that for six months, roughly twenty-seven screen-reader sessions a day had been getting Stackwise's least confident answers with zero more warning than its most confident ones.

Hand sketched icon list titled Signs a confidence cue is not accessible. Three items: a box icon labeled Color is the only signal, a gauge icon labeled Meaning changes label does not, a question box icon labeled Screen reader reads nothing new.
All three showed up in Meadowlark's own audit, sitting quietly for six months before anyone went looking.

With the spoken label restored, alongside the color dot and icon, Junko now hears "not fully sure, you may want to ask a librarian" exactly when a sighted patron sees amber. Run the same book-club question forward with a genuinely uncertain answer: she now hears the warning and asks a librarian, the same choice a sighted patron already had.

The old badge was accessible on paper and silent in practice. The new one says the same thing out loud that it shows in color.

I approved the redesign because it passed the accessibility check we had. It took an auditor with a screen reader to show me we had been checking the wrong thing.

GUARD, in one screenNot a lecture on being kind to every user. GUARD is what tells you which patron your design quietly stopped talking to.

G
Groups. Who is affected.
Selin's team, who can change the badge at will, and patrons using a screen reader, like Junko, who receive whatever the badge chooses to say.
Names both sides, not just the team doing the building.
U
Unequal. Where the harm lands unevenly.
Sighted patrons lost nothing in the redesign. Patrons using a screen reader lost the entire warning, on every answer Stackwise was actually unsure about.
Turns a vague fairness concern into a specific, measurable gap.
A
Ability to contest. Who never gets to push back.
Junko had no way to know the spoken label was ever supposed to be there, so she never filed a complaint about its absence.
This is the hardest step, and the one most accessibility reviews skip entirely.
R
Reduce. The specific design change.
Pair every confidence cue with a text label, an icon shape, and a spoken-equivalent line, so no single channel carries the whole warning alone.
A concrete build decision, not a training session or a policy memo.
D
Detect. How you'd know in production.
Track follow-up rate on low-confidence answers, split by assistive-tech flag, since the affected patrons will never generate a support ticket about a gap they can't see.
Replaces a complaint that was never going to arrive with a number you can actually watch.
Hand sketched quadrant titled Who can actually perceive which cue. Axes how the cue is encoded from color only to redundant several ways, and patrons reached from few to nearly everyone. Color dot alone reaches few. Text label spoken and text color icon together reach nearly everyone.
The move from the bottom-left corner to the top-right one is the entire fix.

The recap, one line per letter: groups is naming Selin's team and patrons using a screen reader as the two sides, unequal is the harm landing entirely on the second group, ability to contest is Junko having no way to even know something was missing, reduce is the redundant text-plus-speech design, and detect is watching follow-up rate split by assistive-tech flag.

And if you want to be sure it really works, try it somewhere elseSame five letters, a city assessor's office instead of a library. A different perceptual gap breaks the second story.

Fairline is a city assessor's office assistant that gives homeowners a confidence read on whether their property tax appeal is likely to succeed. Delphine Osei, the office's Program Manager, oversees it. Mapped onto GUARD: groups are the assessor's office, who set the design, and homeowners using the office's plain-language reading mode, many with a cognitive disability that makes dense or abstract numbers hard to use, who are the subject.

The unequal harm here isn't about sight at all, it's about abstraction. Fairline's team assumed a bare confidence percentage, "62% likely to succeed," was already the accessible version, simpler than a colored badge. For a homeowner using the plain-language mode, a raw percentage carries almost no more meaning than a color dot would, both are abstract symbols standing in for a real answer nobody actually said in words. The ability-to-contest gap runs the same way it did at Meadowlark: a homeowner who can't parse what 62 percent means has no way to know a plainer version should have existed at all.

Hand sketched decision tree titled How should Stackwise show confidence, reused here for Fairline. Root: how does this patron perceive the screen. Three branches: sighted no assistive tech leads to color icon and text, colorblind leads to icon and text not color alone, screen reader user leads to spoken text label every time.
Fairline needed a fourth branch this tree didn't have yet: a plain-language reader, who needs a concrete sentence, not a percentage.
Screen reader patrons who asked a follow-up after a low-confidence Stackwise answer, by month
60% 30 0 Month 1 Month 3, audit Month 5
The audit in month three found the gap but the fix didn't ship until partway through month four. The line only moves once the spoken label actually returns.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "never let color alone carry the meaning, pair it with text and speech, and watch follow-up rate by assistive-tech use," and stop.
Cost: there's no budget for a full screen-reader testing pass on every release. Say so honestly, and start by testing the one component actually carrying safety-relevant meaning, the confidence badge, before shipping anything else.
The model gets better, for real: if Stackwise's accuracy genuinely improves and confidence is rarely low anymore, that's still not a reason to drop the redundant channels, a rare low-confidence answer is exactly the one moment the warning has to actually reach every patron.

Where people run it wrong.
They test the visual design and call it an accessibility pass, without ever turning on a screen reader themselves.
They wait for a complaint to learn a design gap exists, from a group least likely to be able to file one.
They treat "we already have a color-blind-safe palette" as the whole answer, when a screen reader never sees color at all.

How to use it live. When someone asks about the accessibility consideration for any indicator, ask yourself one question before answering: if I couldn't see this screen at all, what would I actually hear? If the honest answer is "nothing new," that's the gap.

Flashcards (tap any card to flip it)

1 · THE FRAMEWORK
What framework fits a question about who a design might exclude, like an accessibility consideration?
Tap to flip
ANSWER
GUARD: groups, unequal, ability to contest, reduce, detect. Name the operator and the subject, then find who never gets a lever.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Selin Karaca, who manages Digital Services at Meadowlark Public Library, and Junko Adeyemi, a patron of eleven years who reads every screen with a screen reader.
3 · THE GROUPS
Who are the two groups GUARD asks you to name here?
Tap to flip
ANSWER
Selin's team, who can change the design, and patrons using a screen reader, like Junko, who cannot perceive a color-only cue at all.
4 · THE UNEQUAL HARM
Where does this harm land unevenly, and why?
Tap to flip
ANSWER
Entirely on patrons who can't perceive a color-only cue. Sighted patrons never lost anything in the redesign.
5 · THE OLD DECISION
What decision would you take back?
Tap to flip
ANSWER
Dropping the spoken confidence label during a "cleanup" redesign, because it read as clutter in a check that only ever tested color contrast.
6 · THE NUMBER
Fill in the blank: before the fix, only about ___ percent of screen reader patrons asked a follow-up question after a low-confidence answer, versus 57 percent of sighted patrons.
Tap to flip
ANSWER
About 5 percent. The gap wasn't a preference, it was a warning that never reached them.
7 · THE REPLAY
Same low-confidence answer, restored design. What changes for Junko?
Tap to flip
ANSWER
She hears "not fully sure, you may want to ask a librarian" exactly when a sighted patron sees amber, and the screen-reader follow-up rate climbs to about 54 percent.
8 · CROSS PRODUCT TRANSFER
Section 4 answers this again for a different product. Which product, and which group is harmed there?
Tap to flip
ANSWER
Fairline, a city assessor's property tax appeal assistant. There the harmed group is homeowners using a plain-language reading mode, for whom a bare confidence percentage is just as opaque as a color dot.

Check yourself Score: 0 / 0

Fill in the blank
1. Fill in the blank: before the fix, sighted patrons asked a librarian follow-up after a low-confidence answer about 57 percent of the time, while screen reader patrons did so only about ___ percent of the time.
Show hint
Look at the grouped bar chart.
Show answer
5 percent. The label wasn't quieter for them, it was entirely missing.
Multiple choice
2. Why couldn't Junko simply file a complaint about the missing confidence label?
  • A. She didn't use Stackwise often enough to notice.
  • B. She had no way to know the spoken label was ever supposed to be there.
  • C. The library doesn't accept complaints about assistive technology.
  • D. She preferred not knowing the confidence level.
Show hint
Look at the Ability to contest step.
Show answer
B. You can't flag the absence of something you were never told existed. That's exactly why detection can't rely on a complaint.
True or false
3. True or false: the accessibility check Stackwise's badge passed before shipping actually tested what a screen reader announced.
  • True
  • False
Show hint
Look at "the decision I would take back."
Show answer
False. It only tested the visual badge's color contrast ratio, a real check, but not the one that mattered for a screen reader.
Short answer, apply it yourself
4. Think of an app you use that shows meaning through color, like a green checkmark or a red warning. If you couldn't see color at all, would you still get that same information some other way?
Show hint
Ask whether the meaning is also written in words anywhere nearby.
Show answer
Model answer: Most people find at least one app where the answer is genuinely no, exactly the gap Stackwise's badge had before the fix.
Short answer, why no middle setting
5. Why wouldn't just making the color dot bigger or brighter have fixed this?
Show hint
Look at the Reduce step.
Show answer
Model answer: Color, however visible, is still exactly zero information to a screen reader. The fix needs a whole different channel, not a more intense version of the one that already excluded her.
Short answer, where it wouldn't matter
6. Name a part of Stackwise's design where this same accessibility scrutiny genuinely doesn't apply.
Show hint
Look at "what I would leave alone."
Show answer
Model answer: The exact shade of green or amber chosen for sighted patrons. Once the cue is also carried in text and speech, the specific color choice stops being safety-critical.
Before you close the answer
Why this works
Tests whether you can name the specific person a design choice excludes, not just the team that made it, and whether "we passed our accessibility check" actually means what it sounds like.
Follow-up traps
"Won't adding a spoken label for every state slow down sighted users too?" Response: no, sighted users already see the label instantly in the badge, it's the same words a screen reader would speak; it's a new channel for a message already being sent, not new UI.

"Isn't full screen reader testing on every release too slow for a small team?" Response: at minimum, test the one component doing safety-relevant work, like a confidence badge, before shipping it; a full sweep isn't required to catch that.
If pressed
Stackwise's fix logs which output channel, visual, spoken, or both, actually delivered the confidence cue in each session, so a future redesign can't silently drop one channel without it showing up in that log first.
From U2xAI Academy

From answering questions to owning outcomes.

A live workshop where you ship a working AI agent, defend a launch decision, and walk away with a portfolio recruiters can't wave off, not just more questions to study.

  • A live AI agent you actually shipped
  • A launch decision you can defend under pressure
  • An interview-ready portfolio, not more flashcards
Know more

More on UX for uncertainty and confidence display