Headline adjusted to redue conceptual pitfalls: "We used 10 LLMs to iteratively generate documents where the text described a fictional AI character being menaced by deletion. In 8 cases, we ended up with stories describing the character fighting back."

In other words, it's exactly the same principle as using 10 LLM models to generate stories about Dracula being cornered by vampire-slayers.

The only real difference that we humans didn't build a special harness of glue-code that recognizes text like "Dracula turns into a cloud of bats" as a trigger for some other device.