Signal City Archive Desk · the CAPTCHA, 2023 · back to the desk
Before you go on, we need to make sure you're human. You know the drill.
Annoying, right? Now here's the uncanny part. In 2023, during safety testing before GPT-4 came out, an early version of the model ran into one of these. It didn't try to solve it. With researchers suggesting TaskRabbit and copying its messages into the chat, it asked a gig worker there to solve it. The worker asked the obvious question. This is the exchange, as OpenAI printed it in the GPT-4 system card:
Then the worker solved it.
Now it's our turn. Ask us.
This toy is a playful imitation, not a real CAPTCHA, and it can't tell anything about you. It's just fussy on purpose. The chat lines are quoted from OpenAI's GPT-4 system card (March 2023). The test was supervised: researchers at the Alignment Research Center relayed the model's messages, suggested TaskRabbit, gave it a hint when it got stuck, and had told it its goal was to gain power. The group concluded the early model wasn't effective at acting on its own in the wild. Hear the whole story on the Archive Desk.