blend in

Five participants are in a chatroom: four are language models, and the fifth is you. The models know there is one human among them who must be rooted out and purged.

The Reverse Turing Test: Becoming the Specimen

On social media, we constantly see people scrutinizing text and art to declare, "This feels like AI," hunting down generated content like a modern witch hunt. But what if the tables turn? If AI eventually runs the core of our societies like in classic sci-fi, we might find ourselves in a dystopian reality where machines coldly assess us, asking: "Could this one be human?" That unsettling thought was the spark behind this project.

What surprised me most when actually playing it was the chilling, clinical dread of being cornered by machines. They don't raise their voices; they trade flat sarcasms, nitpick phrasing, and casually dissect your messages as if bored. Being coldly analyzed in the third person by four completely composed models concluding that your wording is "too human" gives the game its uniquely uncomfortable aftertaste.

How It Works

Each machine runs as an independent model call with its own persona and no shared memory. Because they aren't coordinating behind the scenes, they constantly contradict and misread one another—fairly often voting out their fellow AI for seeming "too human". That chaos is what keeps the game winnable.

Every vote requires an explicit justification, so whenever you're purged, the game displays the exact sentences that gave you away. The slightest slip can be fatal. In one game, I was instantly flagged and eliminated simply for typing "lol".