Quiz apps like Duolingo-style trivia or Kahoot test whether you remember facts. Fact-checking courses teach media literacy with old examples. Nobody trains the specific skill people now need every day: spotting the one fluent, plausible, AI-written falsehood sitting between two true facts. Two Truths and an AI turns that into a daily game, and after every round shows how the lie was built (number swap, name swap, made-up rule), so the skill transfers to real AI answers.
The data it runs on
True statements are drawn from Wikipedia articles (via the free MediaWiki API and the monthly dumps, CC BY-SA, attributed), cross-checked against a second source type such as USGS, NHS or a museum page. An LLM then writes one convincing false statement in the same style, using a named deception technique. Limits: Wikipedia can be wrong or out of date, and a "lie" can accidentally be true. Fallback: each lie is checked by a second model against the article and must be contradicted by it; anything disputed is dropped; rounds carry a "report a problem" button and get human review once flagged. The 20 rounds in this demo were written and checked by hand.
1. The mission
Make a billion people a little harder to fool. AI writes fluent text that sounds right either way; we train the reflex to ask "which part would I need to check?"
2. Why people can't resist it
The puzzle is fair but sneaky. Two facts are true, so you have to really think, not just guess.
Your rating says something about you. "Human polygraph, 1,240" is a badge people want to show.
Fooling the AI is delicious. Reverse mode lets you write a lie that beats the machine, then try it on friends.
3. How it works for real
Source: pick a topic; pull 10 to 20 candidate facts from the Wikipedia article's sentences that have citations.
Select truths: a fast model (Gemini Flash or DeepSeek via OpenRouter) rewrites 2 facts as short plain statements and checks they match the source text.
Write the lie: a frontier LLM (Claude or GPT) writes one false statement in the same voice and length, using a named technique, and states the true version.
Verify: a second model confirms the source contradicts the lie, and both truths are supported. Disagreement drops the round.
Calibrate: early players' answers set each round's "fooled %" and difficulty; your rating moves like a chess Elo against round difficulty.
Reverse mode: the model scores your lie for suspicion signals; friends then try it, always labelled as a game.
4. Cost per use
Pennies per round, and rounds are shared across all players: about $0.01 to generate and verify one round, then free to serve thousands of times. Reverse-mode judging is about $0.001 per lie.
Free: the daily challenge and 3 topic rounds a day. Plus ($2.99 a month): unlimited topics, custom topics from any article, classroom mode, full history of techniques you fall for.
5. Growth loop
The daily result grid is posted to group chats ("4 of 5, Fact-checker"). Reverse mode creates "can you spot my lie?" links that friends open to play. Teachers use classroom mode, so whole classes join at once.
6. Risks and guardrails
Lies escaping: every lie is shown with its correction, and share cards never show a lie without the "one of these is false" frame. Reverse-mode lies are blocked on topics like health advice, elections and real living people.
Accuracy: two-step verification, sources shown, one-tap reports, fast corrections.
Licensing: Wikipedia text is CC BY-SA; statements are rewritten and attributed.
Kids: classroom mode uses curated topics and no free-text sharing.
7. MVP build order
Round JSON format and 50 hand-checked rounds across 5 topics.
Play screen with reveal, technique tags and streak.
Daily challenge with a shareable result grid.
LLM pipeline (source, lie, verify) for new topics, with a review queue.