
Abdelrahman El Sayad · Zenodo (CERN European Organization for Nuclear Research) 2026 · 2026
DOI: 10.5281/zenodo.24238493
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Most talk about AI is about big numbers: how many parameters, how high the benchmark score, how long the context window. This paper argues the opposite: some of the most telling facts about these systems hide in tiny, silly answers. Ask a top AI what animal it would be, and it says octopus. Ask it for a random number, and it says 37. Ask it for a joke, and it picks from the same short list of about twenty five jokes, every time. These are not glitches. They are habits so strong that models built by rival companies, in different countries, on different data, all share them. This paper calls them attractors: answers the systems keep falling into, the way water keeps finding the same drain. We collect the best measured ones, add experiments of our own (an octopus test, a chess piece test, and further identity probes run on November 14, 2026, with predictions recorded in advance), explain the likely causes in plain terms, and argue that silly questions are a real tool: they are cheap, anyone can run them, and they show machinery that benchmark scores hide.
No comments yet — start the discussion below.