When an AI feature makes someone’s task harder

Before calling an AI description an accessibility improvement, check that a person relying on it can make the informed choice the original content supports. “The screen reader read it” is too low a bar, especially when the description is someone’s main way of accessing the information they need.

Imagine a membership website that puts its plan comparison in an image. A customer using a screen reader relies on the site’s AI-generated description to choose a membership for one month. The description gives them two prices: Basic is $20 a month and Flexible is $25 a month. What it leaves out is the line beneath Basic saying that a twelve-month commitment is required, while Flexible can be bought for a single month. The customer has heard accurate prices but missed the fact that decides which plan fits, even though that information was available to someone reading the image.

That is the accessibility problem I want product teams to catch. An AI feature can make content easier to reach while giving someone an incomplete basis for a decision.

Give the tester a decision to make

For this hypothetical comparison, I would ask participants to choose a plan for one month and explain what they would pay and commit to. Let them use their usual screen reader and browser, and watch how they find the comparison, move between the plans and check the terms. Can they identify the annual commitment without a researcher pointing it out, then return to that detail after reading the other plan? If the description and checkout disagree, can they work out why and change their choice? Those observations tell you more than asking whether the description sounds clear.

The task also exposes a basic design question: why is essential pricing information trapped in an image? I would fix the underlying comparison by putting the prices and commitments in properly structured text or a table that people can navigate. W3C’s guidance on images of text recommends actual text where possible because it is more adaptable to people’s reading needs.

An AI summary can still be useful, perhaps by answering a question about the plans or explaining unfamiliar wording. The customer should also be able to reach the actual terms without asking the model to reconstruct them. That alternative needs the same attention as the AI feature: “View original” is a poor escape route if it opens the same inaccessible image.

Find out where the important line disappeared

Suppose the customer reports that the description made Basic sound suitable. Start with what they actually received, keeping the version of the source image, the relevant context sent to the model, the model or configuration identifier available to you, and the resulting description. You can then follow the missing commitment through those records. Was the line absent from the image sent to the model, present but omitted from the generated text, or included somewhere the person could not find?

Those possibilities call for different fixes. Changing a prompt will not repair a source image cropped above the terms, and adding another paragraph may not help someone who cannot navigate the answer. The records give you somewhere concrete to start, although they will not fully explain the model’s behaviour. Collect only what the investigation needs, with appropriate privacy controls.

Check the fix where the failure happened

Return to the one-month membership task with affected users and check it in the product. Can they identify Flexible as the plan that matches their stated need, explain the annual commitment attached to Basic and find the source terms? Reading a corrected answer in a developer’s console tells you little about finding it, navigating it and acting on it with assistive technology.

Also check what the correction costs. If the team fixes the omission by making every answer much longer, the person still needs to move efficiently between the plans. Headings or a structured comparison might make the details easier to revisit; a wall of text may require them to reread the whole answer each time.

This is why W3C recommends combining user evaluation with standards assessment. A standards review checks requirements across the interface, while working with users reveals how the task plays out. Neither a clean automated scan nor one successful participant gives you the whole answer.

What would make me hold the feature back?

In this example, I would hold the AI description back as the primary way to compare plans if it still drops the annual commitment, if users cannot reach the source terms, or if they cannot correct their choice before checkout. A useful result would be more specific than “users liked the AI.” It would say which participants could find both prices and commitments, choose for the one-month scenario and recover from a conflicting answer, alongside the standards issues found and resolved.

That gives the team a basis for its release decision and for testing other plan combinations with people who have relevant access needs. Be clear about the limits of what you tested, and revisit it when the content or feature changes. The customer came to buy one month of membership. The evidence that matters is whether the AI helped them understand the options and complete that task with the same important information available to someone reading the image.

Source: 5 October 2026.