Screen readers can describe images with AI now — so why write alt text?

A JAWS user lands on the campus photo you posted this morning, presses a keystroke, and hears a description of it a few seconds later. Nobody at your office wrote that description. So it's a fair question to ask out loud: if the software can do it, why are you still doing it?

Because what the software produces and what you would have written are two different things. AI description tells a reader what is in the picture. Your alt text tells them why the picture is on the page. That gap is the whole answer, and it hasn't closed.

What the screen readers actually do today

JAWS ships this now. Its Picture Smart AI feature opens with INSERT+SPACEBAR followed by P, and from there the user picks what to describe: C for the image or control they're focused on, F for a file in File Explorer, S for the whole screen, W for the current window. The result appears in the JAWS Results Viewer, and Freedom Scientific says neither it nor its AI partners store or train on the images submitted.

NVDA is heading the same direction more slowly. On-device image description sits under long-term priorities on NV Access's published 2026 roadmap — private, local, no data leaving the machine — and it's appeared in alpha builds, but it isn't in a stable release yet. Add-ons cover the gap in the meantime.

Note what both have in common: the user invokes them. Nothing sweeps your feed and quietly fills in the blanks.

"A man holding a hat" is not your caption

NV Access describes the output of its on-device model with an example of its own: a short description along the lines of "a man holding a hat." That is an honest inventory of what the pixels contain. It is not his name, not the fact that he's your new dean of nursing, and not the fact that the hat is the 1987 championship cap he just donated to the college archive.

This isn't a limitation of one model — it's structural. W3C's alt decision tree puts it plainly: the right text alternative depends on the image's usage, context and content, and it has to be determined by the author. The same bird photo earns one sentence on a parks page and three on an ornithology site. A describer that can only see the image can't know which page it landed on.

The reader has to know to ask

When an image has no alt text, a screen reader announces something like "graphic" or reads out the filename — our walkthrough of how screen readers announce your images has the actual output. From there the reader has to decide that this particular graphic is worth a keystroke and a few seconds of waiting. Scrolling a feed of forty posts, most people won't bother forty times. On-demand description is a floor under a bad experience, and a genuinely useful one. A floor isn't a standard.

Scale matters here too. The 2026 WebAIM Million found 16.2% of home page images missing alt text, affecting 53.1% of the top million home pages — an improvement on the year before, but still an enormous number of images where the reader is doing the work you skipped.

What this actually changes about your job

Less than the headlines suggest, but not nothing. It makes the low-value decisions cheaper and the high-value ones more obvious. Decorative images still get an empty alt attribute — knowing when to write nothing matters more, not less, now that a machine will happily describe your divider swoosh if you don't. Names, titles and relationships still come from you, which is why describing photos of people is the category AI misses hardest. And the numbers in a chart or infographic come from your data, not from a model squinting at the axis labels.

The mental shift worth making is this: you are no longer the only possible source of a description. You are still the only source of the right one.