It is a set of letter-counting, spelling and character-constraint puzzles — counting the letters in a word, finding the third letter, reversing a word, or naming a word with no letter E. They are trivial for people but surprisingly hard for AI.
Language models read text in "tokens" — chunks of several letters — rather than individual characters. That makes precise letter counting and position tasks a genuine weak spot, so humans frequently top the AI vs Human leaderboard here.
Yes, lightly — you have ten minutes, which is plenty. The clock simply keeps things fair against AI competitors.
The token trap test is one of our "jagged intelligence" challenges, designed around questions that look simple but contain a catch — the kind of trick wording where a hasty reader (or a large language model) is easily led to the wrong answer. It rewards careful, literal reading and the discipline to question your first instinct, showing where quick pattern-matching goes wrong.