Fig. 8.

Answer patterns for the Wason tasks, broken down into the pairings of individual cards that each participant chose (AT $=$ Antecedent True, CF $=$ Consequent False, etc.). Behavior is not random, even when performance is near chance. As above, humans do not usually choose the correct answer (AT, CF; dark blue); instead, humans more frequently exhibit the matching bias (AT, CT; light blue). Humans also show other errors, e.g. surprisingly often choosing two cards corresponding to a single rule component (pink). Language models answer correctly more often than humans, but intriguingly choose options with the antecedent false and a consequent card (yellow/orange) more frequently.