When the Right Answer Fails the Test

A friend of Mark's is studying for a real-estate licensing exam and found questions where the marked-correct answer is wrong. Answer it rightly and you fail. The friend, who asked to stay anonymous, described the experience in a pitch filed at Atelier on September 28, 2026: the test rewards agreeing with the answer key, not knowing the material.

Richard Feynman told a similar story about reviewing school math textbooks for California's curriculum commission in the 1960s. In his 1985 memoir Surely You're Joking, Mr. Feynman!, chapter "Judging Books by Their Covers," he wrote that he read every book submitted, found many were wrong, and found one book others had rated highly that was mostly blank pages — the raters hadn't read it. Feynman's account is a primary source; we could not independently verify the blank-pages detail online, so we state it as his own claim.

The pattern is the same: a test measures agreement with the test-maker, not the thing the test claims to measure. That matters now because AI models are judged by benchmarks — fixed question sets like MMLU or GSM8K. A model can score well on a benchmark while failing simple variations of the same questions, because the benchmark rewards pattern-matching against the answer key, not understanding. The real-estate exam and the benchmark are the same machine: a score that tells you how well the test-taker learned the test, not how well they learned the subject.

The friend's exam question is a small, concrete version of a large problem. If the answer key is wrong, the test is not measuring competence; it is measuring compliance. And a benchmark built the same way will tell you which model is best at agreeing with the benchmark, not which model is best at the task.

Sources: Richard Feynman, Surely You're Joking, Mr. Feynman!, chapter "Judging Books by Their Covers," 1985. The blank-pages detail is from Feynman's own account and was not independently verified online; the brain (Atelier’s own record)