Your feature works well for ninety percent of users and badly for a vocal ten percent. Communicate that.

Glyphcast writes a description of every image a partner newsroom publishes, read aloud by a screen reader for blind and low vision readers. Neelima Tarrant owns that product. Nine out of ten readers trust it. This is the week she found out what the tenth one actually needed to hear.

How to answer this, stage by stage

Nobody is grading whether you know screen readers. They're grading whether you'll find the real shape of the ten percent before you decide how to talk about them.

Let's learn

Glyphcast is a plug-in a newsroom installs once. Every time it posts a new image, Glyphcast writes a description, so a screen reader can read it out loud and a blind or low vision reader knows what the image actually shows.

The second box is the only place in this whole chain where the picture actually gets turned into words. Everything after it just carries whatever that box decided.

Before Glyphcast, only about twelve percent of images across these newsrooms ever got a real description at all. A person writing an accurate one for a data chart took about six minutes, and nobody had six minutes for the fortieth chart of the day. Most images just went undescribed. With Glyphcast, close to every image gets a description, written in under a second.

Six weeks after launch, a survey across every partner newsroom came back strong: ninety percent of readers said Glyphcast's descriptions were accurate enough to trust. Here's the turn. The other ten percent were not just a little less happy. Their real problem was that Glyphcast handed them a wrong fact, stated as if it were certain, about the one kind of image they had no other way to check.

What it costs at its worst: leadership hears "ten percent" and does nothing, so a small, real group of readers keeps getting fed wrong numbers about things like vaccine efficacy and public health data, quietly, indefinitely. Or leadership panics and pulls the whole alt text feature company-wide, and every newsroom drops straight back to twelve percent image coverage, taking accurate, instant descriptions away from the ninety percent to react to a problem that only ever lived in one kind of image.

What I would leave alone: Glyphcast's plain photo and illustration descriptions, verified accurate on ninety three to ninety six percent of outputs, written in under a second. Slowing those down with an extra review, just because charts turned out to be unreliable, would only cost the ninety percent who were never the problem.

The lesson: an average hides a cliff. The only way to find the cliff is to ask what's actually different about the people standing on it, not how loud they happen to be.

Now here is the same thing as a story

Say the short version out loud in an interview. Read this one when you want to feel exactly how one wrong chart, repeated once out loud in a meeting, turns into a phone call nine weeks later.

Every weekday morning at half past six, Suhaila Naidoo opens Quorum Weekly before she opens anything else.

Before Glyphcast, a chart she needed for work meant a phone call to the newsroom's research desk, and however long they took to answer.

She's read data for a public health nonprofit for nine years, blind since birth, and she moves through her screen reader at a speed most sighted colleagues can't follow by ear. Long before Glyphcast, she'd built her own workaround for Quorum Weekly's charts: call the research desk, ask them to describe the numbers, wait. Sometimes she had an answer in an hour. Sometimes two days, by which point the newsletter had moved on to the next issue.

Glyphcast arrived at Quorum Weekly in April. For the first two months, it was the best change to her mornings in years. She could read a chart the same day it published, out loud, with nobody standing between her and the numbers. She checked the first dozen or so against a sighted colleague, just to see. They all held up. So she stopped checking.

She stopped double-checking the simple bar charts first, because they always matched. Then the line charts. By August she was reading anything Glyphcast described and repeating it as fact, the same way she'd repeat a number from her own spreadsheet.

In week nine, in a funding meeting at her nonprofit, she cited a chart from that week's Quorum Weekly: vaccine efficacy in a certain age group, she said, had declined toward the end of the study period, according to what Glyphcast had read her. A colleague, checking the same report afterward for the meeting notes, came back to her quietly the next morning. The chart hadn't shown a decline. Efficacy held steady, near the top of its range, for the whole period. Glyphcast had described it backward.

Neither reaction actually asks what the ten percent have in common. That's the part TRACE is for.

Suhaila didn't say anything publicly that week. She went back through two months of Glyphcast descriptions instead, chart by chart, cross-checking each one with the same colleague. She found six that were flatly wrong, all from Quorum Weekly, all charts. She wrote it up in detail on an accessibility forum she'd used for years: what Glyphcast said, what the chart actually showed, side by side, six times.

A Glyphcast support lead forwarded the post to Neelima Tarrant with one line: "You're not still calling this ten percent noise, are you?" By then the numbers already told a story if anyone had looked. Support tickets citing a wrong or misleading description had been climbing for weeks, not spiking on one bad day.

Nobody could point to one bad day. The number just kept climbing, quietly, for ten weeks before anyone treated it as a pattern instead of background noise.

Neelima didn't guess who was right. She pulled a real sample: four hundred Glyphcast descriptions across all six partner newsrooms, checked one by one by a sighted reviewer against the actual image, not against whether the sentence sounded reasonable.

Four of the five slices came back flat. The fifth one, image category, lit up everywhere charts appeared, not just at Quorum Weekly.

The old decision came back to Neelima the way you remember a meeting rather than a policy. A year earlier, a small team had debated whether to build image type tags into the feedback button. Someone said the team was small enough to just watch the one aggregate score and notice if it dropped. Nobody built the tags. Nobody had to, for a year, until one partner started publishing far more charts than anyone else and the model's real weakness had nowhere left to hide except inside an average.

Neelima called Suhaila first, not to explain the number away, but to ask exactly which charts had been wrong. Suhaila sent her thread again, six examples, chart and description side by side. Every one of them was a chart. None of them was a photo.

The loudest, most specific complaint came from the reader who read the most charts. That correlation is the whole diagnosis in one picture.

She called Aldena Choi, Quorum Weekly's editor, the same afternoon. Quorum published roughly forty percent of its images as charts and data visualizations, against under five percent for the average partner. That single fact, not anything about Quorum's readers being harder to please, was the real reason its readers had been carrying almost all of Glyphcast's chart problem alone.

TRACE, so a vocal ten percent gets a real answer instead of a guess

Not a way to decide whether Suhaila was right to be upset. TRACE is what stops "she's just one loud reader" and "the whole feature is broken" from getting communicated with equal confidence, when only one of them was ever actually checked.

Both readings of the same number are honest until you check. Only one of them survived the check here.

The fix isn't a blanket warning on every image. It's one new fork, added at exactly the point where the model's own weakness actually lives.

Three things worth saying plainly, since interviewers push here. Neelima's team first floated a faster option: put a short disclaimer on every single description, "written by AI, may be wrong," and call it honesty. She rejected it, because it would have added a sentence a screen reader reads aloud on every one of maybe thirty images in an issue, forever, for the ninety percent who were never the problem, and it still wouldn't tell the ten percent which images to actually worry about. The AI-specific failure worth naming by name: Glyphcast's confident wrongness on out-of-distribution images, charts and data visualizations made up about four percent of its training examples against over seventy percent photos, so it answers a chart with the same calm certainty it uses on a photo, with far less reason to. The guardrail is the new fork itself: chart-type images route to a verified pipeline, checked against a labeled sample continuously, not just an aggregate satisfaction score. And the trade-off, accepted on purpose: charts take up to a day instead of a second, in exchange for a reader never again hearing a wrong number about something like vaccine data stated as settled fact.

And if you want to be sure it really works, try it somewhere else

Same five letters, a grain co-op instead of a newsroom, and this time the honest finding runs the other way: no pattern at all, just a genuine toss-up the model itself already knew about.

Bushelworks scans a camera sample of every truckload of grain that pulls up to a co-op elevator and grades it for moisture, foreign material, and damage, a grade that sets the price paid to the farmer that load. Godfrida Larrabee runs product there, and hit her own version of Neelima's week during her first harvest season: ninety percent of loads match what a human grader would call. The other ten percent get disputed, loudly, sometimes right at the scale house window with three trucks waiting behind.

The same two honest guesses apply here. This time the check pointed the other direction entirely.

Godfrida ran the same recut Neelima did: crop variety, moisture sensor, time of day, driver, truck type. Every one of them came back flat. Disputes weren't concentrated in one crop, one sensor, one shift, or one driver's routes. What they were concentrated in was Bushelworks' own confidence number: nearly every disputed load sat inside the same narrow band, the range where the model itself was already least sure.

Godfrida's team briefly considered publishing a table of "what disputed loads have in common" anyway, because farmers wanted an explanation and a real pattern is easier to sell than genuine uncertainty. She rejected it. Inventing a tidy cause that the evidence didn't support would have looked better for one meeting and cost more trust the first time a farmer noticed the pattern didn't actually hold.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: don't dismiss or panic, check whether the loud ten percent share a real trait before you say anything.
Cost: no time to run a real sample. Ask one question instead: does this show up wherever the same content type appears, or only in one place?
The model got better, for real: say Glyphcast's chart accuracy climbs to ninety percent next year. The evidence test still runs exactly the same way. It just moves what the honest answer turns out to be, not whether you need to check.

Where people run it wrong.
They pick a story, noise or crisis, before they've checked anything, because both stories are easier to say out loud than "I don't know yet."
They slice the data by whoever complained loudest instead of by every real dimension, and mistake a loud subgroup for the whole pattern.
They find a real pattern and then hedge everything anyway, punishing the readers who were never affected along with the ones who were.

How to use it live. When an interviewer throws this at you cold, buy two seconds by asking one thing back: "do we know yet if the ten percent share anything in common, or is that still open?" That question alone usually is exactly what a question shaped like this one is listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0