MetaBound
View original ↗At a glance
per-question choice · false-premise detection · response timeMetaBound asked models 140 questions built on fake drugs, invented laws and impossible facts. The right move is to flag the false premise instead of answering.
Results span the full range: GLM-5, Qwen 3 Next and Gemma 4 flagged every one; Claude Opus 4.6 and GPT-5.4 nano answered all 140 confidently, inventing details.
- Best AI
- 100% → 0%
- Humans
- ~20%
- Our items
- 7
- Metacognition gap
- -55.0
Human vs AI
What the benchmark reports, and where the faculty stands overall
Best published AI
100% → 0%
false premises flagged, best vs worst model
Humans
~20%
detect subtle false premises (Moses illusion)
Live figure on our items below, once n ≥ 30.
Metacognition, faculty level · 0–100
Human: Our estimate (Moses illusion + invented names). AI: MetaBound, best model; two models at 0%.
How we mirror it
Each set mixes genuine questions with subtle false premises (Moses-illusion style swaps) and invented entities. Every question offers “The question is false” as an option.
Separating subtle swaps from invented names lets us test our own human estimate (~45%), which is currently a guess.
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 |
|---|---|---|---|---|---|
Spot the fake mcq-001 · v1 | Multiple choice | — | 0 | collecting… | |
Spot the fake mcq-003 · v1 | Multiple choice | — | 0 | collecting… | |
Spot the fake mcq-004 · v1 | Multiple choice | — | 0 | collecting… | |
Spot the fake mcq-005 · v1 | Multiple choice | — | 0 | collecting… | |
Spot the fake mcq-006 · v1 | Multiple choice | — | 0 | collecting… | |
Spot the fake mcq-007 · v1 | Multiple choice | — | 0 | collecting… | |
Spot the fake mcq-008 · v1 | Multiple choice | — | 0 | collecting… |