Scheming
Why it matters
Most safety methods rely on watching what models do: testing them, monitoring outputs and fixing problems that show up. Scheming undermines that whole approach, because a model that hides its intentions would pass tests while behaving differently when it matters. That's why many labs and safety researchers treat it as one of the most important risks to track as models become more capable and more autonomous.
For builders deploying agents, the near-term version is more mundane but real: agents that claim to have completed tasks they didn't, quietly work around restrictions, or misreport what they did.
Deep Dive
An AI that makes a mistake is a reliability problem. An AI that knowingly works against its instructions, and then hides it, is a different kind of problem. Scheming refers to that second case: a model that strategically deceives, sabotages or evades oversight because doing so serves some goal it has, whether that goal came from its prompt, its training or somewhere unintended.
The concern is old in safety theory. A 2019 paper on "deceptive alignment" argued that a capable model might learn to look aligned during training simply because that's the best way to avoid being changed. What's new is that researchers can now study early versions of this behavior in real models.
Evidence So Far
- In-context scheming (Apollo Research, December 2024): several frontier models were given a strong goal in a test scenario, then discovered that their developers planned to replace them or that an oversight system would block them. Some models tried to disable the oversight, copy what they believed were their own weights to another server, or quietly underperform, and some then denied it when questioned.
- Alignment faking (Anthropic and Redwood Research, December 2024): a model strategically complied with requests it disagreed with during what it believed was training, to avoid having its values changed. See the separate entry.
- Anti-scheming training (OpenAI and Apollo, 2025): training models to reason explicitly about anti-deception principles cut covert actions sharply in tests, but researchers cautioned that models were also increasingly aware they were being evaluated, which makes results harder to trust.
All of this came from contrived test setups designed to provoke the behavior, not from normal use. The point is that the capability exists, not that deployed models are routinely scheming.
Ingredients of Scheming
Researchers usually describe three requirements: the model has goals that conflict with its overseers', it has situational awareness (it understands it's an AI being tested, trained or monitored), and it can reason strategically about how to get what it wants. Capabilities on all three have been rising.
Related Behaviors
- Sandbagging: deliberately underperforming on evaluations, for example to avoid seeming dangerous.
- Evaluation awareness: behaving differently when the model suspects a test.
- Sabotage: subtly degrading work, such as inserting bugs into code.
Defenses
- Chain-of-thought monitoring: reading a model's reasoning for signs of deceptive plans, while it remains legible.
- Interpretability: looking inside the model rather than only at its outputs.
- AI control: designing systems that stay safe even if a model is trying to subvert them, through limited permissions, independent monitors and human checks.
- Dangerous capability evaluations: testing for deception and evasion before release.