Spot the hallucination
A round of AI-generated claims and citations flashes up — most accurate and grounded, one or two subtly fabricated: a plausible but nonexistent citation, a confidently wrong number. Tap the fakes against the clock. Each round ends by revealing exactly what was faked, the tell you could have noticed, and the one check that would have caught it. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Read the claims — they all sound equally confident and authoritative.
- Tap the ones you think are fabricated, then submit.
- See what was really fake, the tell in each, and the single check that would have confirmed or debunked it.
What this clears up (the fundamentals)
- Fluent confidence is not a truth signal — a hallucination reads exactly like a fact: same tone, same certainty. The model is equally confident when right and when inventing.
- Fabrications cluster around specifics — precise numbers, named studies, and citations are where models most often make things up, because a plausible-sounding specific is easy to generate and hard to spot.
- The defence is a habit, not a vibe — you can't feel your way to accuracy. For any load-bearing number, name, or citation, run the one check that confirms it before relying on it.
- Missing one feels different from being told — the point of the timer and the reveal is to make the miss felt, so the reflex sticks.
Where it's used
An AI-literacy reflex trainer — the fastest way to internalise that confident output needs verification, and to build the habit of checking specifics. It's a miatz build-lab concept playable; learn to build it yourself.
FAQ
What is an AI hallucination?
Content a model generates that is fluent and confident but factually wrong or entirely invented — a fake citation, a made-up statistic, a confidently incorrect answer. It isn't lying; it's producing plausible text without a ground-truth check.
Why are fabrications so hard to spot?
Because they're generated to be plausible and are delivered in the same authoritative tone as true statements. There's no stylistic tell — the only reliable signal is external verification.
What's the "one check"?
The single fastest way to confirm a specific claim: search for the citation, look up the number, or check the official source. The game shows the exact check for each fabrication after you submit.
Is anything uploaded?
No. The claims and scoring run entirely in your browser — nothing is transmitted or stored. Turn off your Wi-Fi and it still works.
Limits
A teaching model with a fixed bank of grounded and fabricated claims — real hallucinations are open-ended and context-dependent. The reflex it trains (verify specifics, distrust tone) transfers directly.
Related
Part of the miatz playgrounds. Explore the rest from the playground catalog.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.