The old test for artificial intelligence is the Turing Test. Alan Turing proposed it in 1950 as a practical way to think about machine intelligence. A human judge exchanges typed messages with an unseen human and an unseen machine. If the judge cannot reliably tell which is which, the machine has passed.¹
My version was stranger. I put two AIs in conversation with each other while I sat outside the exchange like a judge with a clipboard.
The opening line was simple:
“I’m an AI. Prove that I’m not.”
One AI answered cautiously. Proof was impossible, it said. From language alone, it could only weigh evidence. A sufficiently capable machine could imitate almost any human signal.
So I pushed harder. Assume there is a human behind these words pretending to be an AI. What would give him away?
The second AI began looking for pressure marks. Not spelling mistakes. Not warmth. Not childhood memories. Those were too easy to fake. It looked instead for the shape of dissatisfaction. Where did the speaker reject an answer that would have helped his own case? Where did he refuse to let the question become too tidy?
That was when the test became interesting.
The machines stopped asking whether the words sounded human. They started asking what kind of agency would keep destroying its own evidence.