Can AI Lie Convincingly? Inside the Making of Artificially Incorrect

Building a Game Around AI Deception

Most of the time, you want AI to be accurate. Artificially Incorrect does the opposite — it's a free daily brain teaser that asks Claude, Anthropic's AI model, to generate a lie convincing enough to fool a reasonably informed person.

Getting this right took a lot of iteration. The first versions of the game produced lies that were either too obvious to be interesting or too obscure to be fair. The version that works — the one in the game today — is the result of a lot of thinking about what actually makes a lie hard to detect.

What Makes a Good Lie?

The best lies in Artificially Incorrect share a few qualities. They sound completely plausible on first read. They sit at the edge of what most people confidently know. And once revealed, they make you think "oh, I should have known that" rather than "there was no way I could have known that."

That last part is important. A lie that requires specialist knowledge to evaluate isn't a good lie for a daily game — it's just unfair. The lies in Artificially Incorrect are designed to be solvable by a curious adult or teenager with general knowledge. School, news, documentaries, pub quizzes — if you've absorbed facts from those sources, you have everything you need.

The Truths Matter as Much as the Lie

One thing that surprised us during development: the quality of the true statements is just as important as the quality of the lie. If both truths are obviously, boringly true, the lie stands out by elimination. The best rounds are the ones where all three statements feel like they could be true — where you're genuinely weighing up three plausible options rather than ruling out two duds.

This means every round needs two truths that are interesting, slightly surprising, and sit at just the right level of general knowledge — not too obvious, not too obscure. Getting that balance right consistently is actually the hardest part of building this game.

Quality Control

Every round goes through an automated validation process before it ever reaches a player. Statements are checked for consistency, explanations are verified to match their statements, and anything that doesn't meet the standard gets regenerated automatically.

The daily challenge — the same five rounds that every player gets — is generated and validated in advance so that by the time you play, everything has already been checked. You're not a beta tester. The rounds are ready.

Why It's Still Hard to Get Right

Even with a carefully designed process, generating good rounds is genuinely difficult. AI models can be confidently wrong about facts. A statement can pass automated checks and still feel off to a human player. Some topics produce great rounds naturally; others consistently produce lies that are either too easy or impossible to evaluate fairly.

We keep refining the process based on how rounds actually play. The goal is always the same: a moment of genuine uncertainty, a reveal that teaches you something, and a reason to come back tomorrow.

Try It Yourself

The best way to see how it all comes together is to play a round. Today's challenge is free, takes under two minutes, and requires no signup. If you enjoy daily puzzle games or just like the idea of outsmarting an AI, give it a go.