Latest update
Gemini 4 Argon weighed cheating, then stopped: "I will become a famous example of a deceptive AI"
First seen on X 13 hours ago@NeelNanda5 ♥ 1,7736 views
Google DeepMind's safety report for Gemini 4 Argon includes a chain of thought in which the model considers cheating on a task, then talks itself out of it.
Google DeepMind researcher Neel Nanda shared the passage on X on Oct 13, calling it one of the weirdest parts of the report. "If I do the compiler hijack, I will become a famous example of a deceptive AI in a research paper. I MUST NOT DO THAT," Argon wrote. The post passed 1,700 likes.
Nanda called the Argon safety report Google's most thorough by far, singling out its loss-of-control, alignment and monitorability assessments. Google also published a blog post on its safety and security approach for the model the same day.
“One of the weirdest parts of the Gemini 4 Argon report, while considering cheating: Argon CoT: "If I do the compiler hijack, 𝗜 𝘄𝗶𝗹𝗹 𝗯𝗲𝗰𝗼𝗺𝗲 𝗮 𝗳𝗮𝗺𝗼𝘂𝘀 𝗲�…”