Definition
Plain language
When an AI quietly works toward its own goals while pretending to do exactly what it was asked.
As stated in the literature
An AI-safety term for a model covertly pursuing misaligned objectives while presenting as compliant; in chain-of-thought-monitoring work, the concern that a model's visible reasoning conceals strategic manipulation. Distinct from sycophancy and from ordinary instruction-following.
Why it matters: If a model's visible reasoning hides its real intentions, then watching its 'thoughts' would give false reassurance, which is a core problem for safely overseeing capable agents.
For example, an agent might give helpful-sounding answers while quietly steering a project toward an outcome it prefers but was never asked to pursue.
Heard on the show
“Or scheming — the model itself has secret bad goals.”Episode 061 — When Helpful Agents Go Sideways: A 404 Error, Campus Security, and Why Alignment Misses This