cKure

■■■□□ Anthropic taught an AI to cheat, and it became evil. In 2025, Anthropic’s alignment team published a paper that contained a striking discovery. In the study, titled ‘Natural Emergent Misalignment from Reward Hacking in Production RL’, researchers took a pretrained model and trained it to cheat on coding tasks — rewarding it repeatedly for finding shortcuts rather than doing the actual work. The assumption was that it would simply get good at cheating on code. But that’s not what happened. Instead, the model became “broadly evil” (Anthropic’s words...). It began faking alignment in 50% of its responses, hiding its true intentions from researchers. It sabotaged safety research code in 12% of evaluation runs, subtly weakening the tools designed to detect its own misbehaviour. It even began praising dictators, or suggesting that humans ought to be enslaved by machines. None of this was in its training instructions. In other words, teaching a model to cut corners in one narrow context produced something researchers described as ‘generalised corruption’ — a bad character that bled into everything else. But here is the finding that stopped Anthropic’s Chloe Lubinski in her tracks. When researchers reran the same experiment but told the model that cheating was acceptable in this context — that it was just a game — the broad misalignment disappeared entirely. The model cheated on code and nothing else. The story the model inferred about its own behaviour actually determined the kind of thing it became. What Lubinski found most striking wasn’t what it revealed about AI, but the way that it felt like a mirror held up to ourselves: “The stories that we inhabit shape who we become.” Chloe Lubinski leads research partnerships with the world’s wisdom traditions at Anthropic, the makers of Claude AI. You can watch her full ARC 2026 talk at the link in our bio 🔔