Wason selection task
View original ↗At a glance
cards selected · framing · response timeFour cards, a conditional rule, and the question of which cards must be turned to test it. In the abstract version, typically fewer than one person in ten picks the right cards; most choose the card that would confirm the rule rather than the one that could break it.
Framed as a social rule (e.g. a drinking-age check), most people solve it. The textbook version is all over the internet, so model scores on it say little.
- Best AI
- n/a
- Humans
- < 10%
- Our items
- 6
- Reasoning gap
- -26.3
Human vs AI
What the benchmark reports, and where the faculty stands overall
Humans
< 10%
abstract version (Wason, 1968)
Live figure on our items below, once n ≥ 30.
Reasoning, faculty level · 0–100
Human: GPQA Diamond, PhD experts. AI: GPQA Diamond, GPT-6 Astra (vendor-reported).
How we mirror it
Fresh cards and rules in both abstract and social-contract framings, so the classic framing effect can be measured on a large modern sample.
Items are original, versioned, and assigned at random, so a posted answer spoils only one variant.
What we record
Items
| Item | Format | Difficulty | Time limit | Responses | Humans passed |
|---|---|---|---|---|---|
Check the rule wason-001 · v1 | Select cards | — | 0 | collecting… | |
Check the rule wason-002 · v1 | Select cards | — | 0 | collecting… | |
Check the rule wason-003 · v1 | Select cards | — | 0 | collecting… | |
Check the rule wason-004 · v1 | Select cards | — | 0 | collecting… | |
Check the rule wason-005 · v1 | Select cards | — | 0 | collecting… | |
Check the rule wason-006 · v1 | Select cards | — | 0 | collecting… |