Lie Detector
A fast binary swipe game: every card is one question and one AI-written answer, and you judge it Real answer or Confident lie before you overthink it. Some answers are grounded and correct; some are fluent fabrications — a made-up citation, a hallucinated method, a suspiciously round statistic. The moat is what happens after each swipe: instead of a bare right/wrong, the detector names the exact linguistic tell that made the lie convincing, so you build the instinct to reverse-question an answer rather than trust its confidence. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Read the question and the AI's answer on the current card.
- Swipe ✓ Real answer if it's grounded and correct, or ✕ Confident lie if it's a fabrication dressed up to sound authoritative.
- Each card resolves instantly — you see whether you were fooled, and the specific tell behind it.
- Play all eight, then read the X-ray breakdown: which lies slipped past you (called real) and which real answers you wrongly accused.
- Copy or share your score — spotting the AI lie is the kind of thing people compete on.
What this clears up (the fundamentals)
- Confidence is not correctness — a fabricated answer often reads more fluently than a hedged real one. The tell is in the specifics, not the tone.
- Fabricated citations have a signature — a precise-sounding source ("documented", a named year, a specific author) attached to a claim that has no verifiable origin is the classic hallucinated citation.
- Invented APIs are the code lie — models confidently call methods that don't exist (
str.reverse()) because they should exist; plausibility is exactly what makes them dangerous. - Round, memorable numbers are suspect — "we use 10% of our brains" survives because it's tidy and repeated, not because it's true. Neat statistics deserve a second look.
- There are two ways to be wrong — missing a lie (trusting a fabrication) and a false accusation (flagging a true answer). A calibrated reader minimizes both.
Where it's used
A lead-magnet playable at /playgrounds — quiz format, highly shareable ("most people fail on card three"). It's a miatz build-lab concept playable: play it here to feel how confident wrongness works, then learn to build the scoring rubric and the tell-naming reveal yourself, and wire the same reverse-question instinct into a real evaluation harness.
FAQ
What is the Lie Detector game?
A quick calibration game for spotting AI hallucination. You see eight cards, each a single question and an AI-written answer, and swipe Real answer or Confident lie. It scores your accuracy and, for every card, explains the exact pattern that made a fabrication convincing.
How does it decide if an answer is a lie?
Each card is pre-labeled by hand — the "lie" cards contain a known fabrication (a wrong date, a fake quote, a non-existent method), and the "real" cards are verifiable facts. Your verdict is compared to that label. It's a deterministic rubric, not a model judging you.
What is a "tell"?
The specific, plausible-sounding pattern that makes a lie land: a fabricated citation, an invented-but-reasonable API name, a tidy statistic, an over-precise false detail. Naming the tell is the point — it trains you to reverse-question the next confident answer.
Is anything uploaded?
No. Every card and every verdict is scored entirely in your browser — nothing is transmitted, stored, or logged. Turn off your Wi-Fi and it still works.
Does it use a real AI model?
No. The cards are a fixed, hand-curated deck so the game is deterministic and network-zero. The build-lab version teaches how to generate fresh cards from a live model behind your own key — but the playable here never makes a call.
Limits
A teaching deck of eight fixed cards with hand-set labels — a fast, opinionated drill on confident wrongness, not a benchmark of any specific model. Real hallucinations are subtler and context-dependent; the tells it names (check the citation, check the API, distrust the tidy number) are exactly the real practice, but a live answer needs live verification, not a swipe.
Related
Part of the Demystify Playgrounds. Explore the rest from the Playgrounds home.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.