Who carries the error rate

In June, Aktion Mensch published a study on how people with disabilities in Germany use AI. Its headline conclusion is that AI can strengthen participation, under certain conditions. I read the 47-page PDF for a number that wasn't in the headline.

What the survey is

Ipsos asked 574 people in December 2025, all from an online panel of people with impairments. The study calls it "eine nicht quotierte Stichprobe ohne nachträgliche Gewichtung." Not representative. Not weighted. And everyone in it already has internet access, so the 25 percent who have never used AI is a share of people who are online. Among people who aren't, it can only be higher.

Of the 574, 16 percent use AI daily, 26 percent several times a week, 31 percent weekly or less. The 417 users mostly ask it for facts and answers (73 percent) and advice (48 percent). About a third use it to make communication easier, with translation, subtitles or read-aloud.

(The press release turns this into "73 percent of AI users use it regularly." The study says 73 percent of all respondents use it at all. I went with the study.)

The question about checking

One item asked whether people can tell when the AI is wrong: "Ich kann herausfinden, ob KI-Anwendungen richtige Antworten geben oder Fehler machen." 58 percent of users agree. People with sensory impairments rate themselves "kompetenter in der Beurteilung der Korrektheit von KI-Antworten" than the other groups.

That is a self-assessment. The study measured no actual errors. A diary study from CHI 2026 did, with a different question in a different country: 20 blind and low-vision people used an AI app for two weeks, rated its answers as trustworthy, 3.76 out of 5 on average, and the AI was wrong 22.2 percent of the time. The two studies don't contradict each other. They show the distance between believing you can catch an error and catching it.

Imagine the same wrong answer twice

Imagine two blind people looking at the same product photo. The first asks the AI on their own phone what is in it. The second reads the alt text the shop generated with an AI and never reviewed. Both get the same wrong answer.

The first one knows a model answered. They can ask again, ask differently, or ask a person. The second one doesn't know a model was involved at all. Same error rate, different person holding it.

The first kind is common and it's fine. In this study, about a third of the people who use AI use it to make communication easier. The second kind is the one I wrote about in June: unreviewed generated alt text, where the user carries the error and nobody told them.

What nobody asked

The questionnaire has no item about errors people actually ran into, none about image descriptions specifically, and the word "Datenschutz" doesn't appear in it once. Those three would be the next questions, and the first would show whether the 58 percent holds up.

It did ask whether the AI apps themselves are accessible. 79 percent of users say they can use them easily and without barriers. Among non-users, 28 percent say they can't and 41 percent don't know, and the study counts both as a reason people stay out. Among users with cognitive impairments, about a quarter say no. The group is small, and the study calls it a trend, not a finding.

A demand without a law

87 percent of all respondents say there have to be rules that make AI applications accessible by obligation. As far as I can read the EU's rules, that law doesn't exist in general form. The AI Act requires the "you are talking to an AI" notice to meet the applicable accessibility requirements, and the Commission says that adds none of its own. The European Accessibility Act covers AI only where it is part of a covered service, a bank's chatbot or a shop's assistant. A general-purpose AI app on someone's phone is neither.

A survey can ask people whether they think they can catch an error. A product team that puts AI between a disabled user and their information decides whether that user will have to.