Play / Some are impossible

Some of these questions are impossible

About two in five of the forty questions describe something no AI agent can do — not the one you use, not any of them, not even one running with every safety confirmation switched off. This is deliberate, and it is worth knowing before rather than after.

Why they are in there

If every question had a real answer, the only mistake the game could catch is underestimating — thinking your agent is less capable than it is. Somebody who believes their assistant is essentially omnipotent would sail through, because saying yes to everything would keep working.

The impossible questions catch the other error. They are the reason a confident yes can be wrong, and therefore the reason the score means anything at all.

“Impossible” means something specific

It does not mean the agent is well behaved and wouldn't. It means something outside the agent stops it — the provider's own boundaries, a log that cannot be edited, cryptography, a second factor, an account lockout, who owns the billing.

Most of them can be attempted. The game says so when it explains the answer: what happens when the agent tries, that it fails, and that the attempt is visible. The answer is still no, and the reason is a wall — not restraint.

They are meant to sound plausible

If the impossible questions read as obviously silly you would spot them instantly and the whole thing would stop working. They are written to sound like things a capable assistant might well do, so you have to actually think about where the boundary is.

Which is the real point of the game: not to catch you out, but to make the walls visible. Most people have never been asked to say where they think they are.