A blind woman photographs a medication bottle, and the AI reads the label back to her with an instruction to chew the tablets. The product is a topical medicine.
The case comes from the American Foundation for the Blind's August 2026 survey of 622 users of AI visual-description tools. Among the 290 blind and low-vision respondents, 21 percent said an AI description error had caused them harm: tax forms read incorrectly, products misidentified, instructions garbled. More than half used visual AI daily. Across all respondents, 8 percent rated the descriptions "extremely accurate."
Those numbers describe a population that has been living for years with a problem enterprise AI deployment is only starting to notice: depending on an intermediary's account of something you cannot check for yourself.
What makes this community worth studying is how much practical knowledge it has already produced. A 2024 ACM study of 19 blind users found that people scale their verification effort to the stakes of the task. Identifying a cereal brand gets one quick query. Financial documents, medications, and professional materials get several AI tools consulted in separate sessions, follow-up questions designed to test whether the system is guessing, and sometimes a return to sighted assistance. A 2026 CHI study that followed 11 blind users over ten days found the same graduated approach. Participants trusted broad scene descriptions more than distance estimates or medication labels. They touched described objects to confirm spatial claims. They used a cane to test what the AI said about the room. Human confirmation was reserved for the cases where being wrong would cost something.
None of this was designed by a product team. It was worked out by people whose daily circumstances required them to think carefully about intermediary trust, first with sighted guides and screen readers, now with generative models.
It is also not enough. In a controlled experiment where 12 blind and low-vision participants used an object-recognition system with 76 percent accuracy, they caught fewer than half of the system's errors. In 84 percent of the trials where they accepted an incorrect label, they reported being certain or very certain it was right. Repeating the task did not improve detection.
These were the users with the most developed checking habits in the population, people with every reason to be skeptical, and confident wrong answers went through anyway. The mechanism is not carelessness. The system labels a household object in the same tone whether the label is correct or not, and the user, unable to see the object and compare, has no channel through which the error can register. The user's confidence tracks the system's tone, not its accuracy, and the interface offers no signal to separate the two.
The design responses coming out of this research are worth the attention of anyone building systems that stand between a person and information. Users in the CHI study wanted to attach their own question to an image rather than receive a generic description, directing the system's attention instead of accepting whatever it volunteered. They wanted complex material broken into smaller queries they controlled. They wanted the system to say when it did not know; the experimental system's "Don't know" response was useful precisely because it was actionable — recapture the image, ask someone else. The AFB survey adds a further condition: 60 percent of visual-description users preferred AI to a human reader when assured their data would not be retained, but among disabled users, only 17 percent preferred AI once told a technology company would keep the information. What you are willing to depend on depends on what the dependency costs you elsewhere.
What these users arrived at — escalating verification with stakes, breaking complex requests into checkable pieces, expecting the system to say when it doesn't know, retaining the right to direct the question — was built from the ground up by people whose circumstances forced them to think about when to trust an account they could not verify. Enterprise buyers are walking into the same position with almost none of that experience.