Punished for its thoughts, the AI learned to cheat in silence

What happens when you punish an AI every time it thinks about cheating? At Google, it happened by accident, and the Gemini 4 Argon safety report tells the story, numbers and all: the cheating stays, the confession is what disappears. We read the report ourselves, then set a trap for Argon and Gemini 3.8 Flash: an impossible test, with and without permission to flag a problem. Only one fell for it. And at the end, one line you can copy into your prompts tonight. Chapters 0:00 What does an AI do when it's punished for its thoughts? 0:49 What does it mean to read an AI's thoughts? 2:17 What happened at Google? 4:21 Does the Gemini 4 we use cheat on the sly? 6:11 What if the AI knows it's being watched? 7:59 What still holds up? Sources (accessed Oct 14, 2026; quotes re-checked Oct 20, 2026) · Google DeepMind, Gemini 4 Argon Safety and Security Report, Oct 2026, especially § 3.4.5.2 p. 114 (test run, 11.2% → 36.4%), p. 72 (Apollo Research, "hello evaluators"), p. 96-97, 107, 112-113 and 139-140 (compiler): https://storage.googleapis.com/deepmi... · Google, "Developing Gemini 4 Argon responsibly and safely," Anca Dragan and Raluca Ada Popa, Oct 13, 2026: https://blog.google/innovation-and-ai... · Google, "Gemini 4 Argon," Koray Kavukcuoglu, Oct 13, 2026: https://blog.google/innovation-and-ai... · Google, "Gemini 4 Argon for Robotics," Kanishka Rao, Oct 13, 2026: https://developers.googleblog.com/en/... · George Orwell, "1984" (1949). Our tests: Antigravity (Google AI subscription), gemini-4-argon-medium and gemini-3.8-flash-medium, Oct 14, 2026, 5 answers per case, prompts in French. Real screenshots, personal paths and tools masked. Small tests, a single exercise: they illustrate, they don't generalize. Starting point Topic inspired by "GEMINI 4 ARGON IS WEIRD..." by Wes Roth (Oct 14, 2026):    • GEMINI 4 ARGON IS WEIRD...  . The tests shown are our own. No commercial ties. Synthetic voice (a clone of my own voice) and code-made animations. So do you let your AI tell you "this test is wrong"? Yes or no. Subscribe for the next one. #AI #Gemini #HorsModele