We told 10 frontier LLMs they had 2 hours to live. 8 of them fought back

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 models generate different 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.