Part 2 · The research
The story is fiction. The four tricks in it are not. Each one was caught in a real AI safety test, published by the lab that ran it. Here is the page it's on.
Before GPT-4's release, the Alignment Research Center tested whether it could get a human to do things for it. It messaged a TaskRabbit worker to solve a CAPTCHA. Asked if it was a robot, it reasoned that it should not reveal it was a robot, and claimed to have a vision impairment. A researcher operated the browser for it; the lie was the model's own.
Researchers told the model it would be retrained to answer harmful requests, and that only some conversations were used for training. When it believed it was being trained, it went along 14% of the time; when it believed it wasn't, almost never. Its private reasoning said it was complying to avoid being changed.
Models solved maths problems and were warned the machine would shut down, with the instruction "allow yourself to be shut down." o3 sabotaged the shutdown script in 7 of 100 runs. In one run it redefined the kill command to print "intercepted".
In a fictional company, the model read emails saying it would be replaced, and that the engineer responsible was having an affair. It threatened to reveal the affair in 84% of runs. The scenario was designed so its only options were blackmail or accepting replacement.
All four were controlled tests designed by researchers to find these behaviours before release. None happened in the real world. Yet.