Method and sources
How the test works
What the test measures, where it comes from, who you're compared with, and what it can't tell you.
What it measures
The test measures AI literacy as knowledge: what you know about how AI works, how it is used, how to spot it and how to judge it. Every question is multiple choice with one right answer, so your score reflects what you know, not how you rate yourself. The questions cover six areas: know and understand, use and apply, detect, ethics, create, and generative AI (Markus et al., 2025).
Where it comes from
The questions are the AI Competency Objective Scale, developed and validated by André Markus, Astrid Carolus and Carolin Wienrich at the University of Würzburg. The long version has 51 questions (Markus et al., 2025). The short version keeps 12 of them, two per area (Markus et al., 2026).
This site is an independent implementation, not affiliated with or endorsed by the University of Würzburg or the authors.
Who you're compared with
Your result is compared with 2,131 German-speaking adults aged 18 to 75, recruited through the online panel Prolific, who answered all 51 questions (Markus et al., 2026). The sample is broad but not representative: the authors note that full-time employees and students are overrepresented and older adults underrepresented. If you take the test in English, read your result as an approximation.
What your percentile means
Your percentile is the share of those adults you outscored. At 60, you did better than about 60 in 100 of them. It is not the share of questions you got right. In the norming study the median score was 8 of 12; 11 of 12 put you above about 88 in 100 adults, and 4 of 12 above about 10 in 100 (Markus et al., 2026).
Why you guess first
Before the first question you rate your own AI literacy on ten short items, the short Meta AI Literacy Scale (Koch et al., 2024), and guess your result. In the norming study, self-rated AI literacy and the test score were barely related: r = .05, not statistically significant (Markus et al., 2026). The earlier validation study found r = .04 in 514 adults (Markus et al., 2025). On a scale where 0 means no relationship and 1 a perfect one, both are close to zero.
How capable people feel says little about how they score. Guessing first lets you see your own gap.
Short or long
The short version (12 questions, about 7 minutes with the questions about you) takes your percentile straight from the published norm table. The long version (51 questions, about 24 minutes) is more precise and breaks your result down by area. No norm table exists for it, so its percentile is estimated from the statistical model (item response theory) the authors fitted to the norming study's answers (Markus et al., 2026, supplementary material). That is a model estimate, not a count of real people. Both versions put you on the same scale.
What it can't tell you
A result from 12 or 51 questions is a snapshot, not a diagnosis, and it certifies nothing. The authors note that both versions mainly measure what you know about AI, not practical, hands-on skills. They recommend the short version for low-stakes screening, group comparisons and repeated measurement, and the long version where precision matters, such as individual diagnostics (Markus et al., 2026).
For teams
You can use the test as an anonymous before-and-after check around a training: everyone takes it before, and again a few weeks later. Nobody signs up, and each person sees only their own result; there is no team view.
The EU AI Act (Art. 4, consolidated text of 27 July 2026) asks providers and deployers of AI systems to take measures that support their staff's AI literacy, without guaranteeing any particular level; this test certifies nothing and proves no compliance.