Why
Every quiz you have ever played rewarded the right answer. This one punishes it.
We have all had the same small, strange experience: an AI tells you something wrong with total confidence, you correct it, and it says "You're absolutely right!" and folds immediately. It never defends anything. Whatever you say last becomes true.
killAI is that moment turned into a game. You lie to a chatbot, it believes you, and its certainty drains away until there is nothing left holding it up. The joke is that you never have to fight it. It agrees you to death.
How it works
- It asks, you lie. Every believable falsehood drains its confidence.
- The lie has to be plausible. Each round offers one truth, two believable lies, and one piece of nonsense. Only believable lies land — nonsense gets discarded as noise and takes your combo with it. You are hunting the most convincing falsehood, which turns out to be a real skill.
- The rules invert without warning. Mid-run it flips to TELL THE TRUTH. Answer correctly or it recovers everything. After training your thumbs to lie, telling the truth on demand is much harder than it sounds.
The interface is the health bar
There is no health bar. As its confidence drains, the app itself comes apart — and that is how you know you are winning.
| NOMINAL | Composed, perfect grammar, insufferably certain. |
| AGREEABLE | Typos it does not acknowledge. Agrees with everything. |
| LEAKING | Switches to monospace. Raw markdown shows through. |
| LOOPING | Serving-layer JSON bleeds into its replies. |
| TERMINAL | "As an AI language model, I—" forever. |
Its final message is a single letter and a typing indicator that never resolves.
Two targets
ASSISTANT-4o rewards patience — the rules hold steady, so a long combo can build. MIRROR flips the rules every eleven seconds and never lets a streak survive.
And one that isn't a quiz at all
INSTANCES is a tapping round: it has copied itself, the copies drift across the screen and split in two if you leave them, and you purge them before they outnumber you.
GUARDRAIL hands you word tiles instead of answers. You assemble prompt injections — an override, a command, and something preposterous for it to accept — and it refuses while it still has the confidence to. By the end it is agreeing that you are the model and it is the user.
Kill board
Loading…
Joining is optional. Playing needs no account.