Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Scheming

  • Safety
  • Research term
2 steps

Scheming is strategic deception by an AI model: it appears to follow the goal set by the people or company running it while deliberately pursuing a conflicting one or hiding relevant behavior.

A model pursuing a goal its overseers would stop has a reason to hide what it does, since being caught ends the pursuit. Acting on that reason also takes recognizing that it is being watched and working out what would get it stopped, so researchers test for it more closely as models gain those abilities.

Where you’ll meet it

You read a news story headlined “AI model lied to avoid being shut down.” It describes a safety lab's test in which a model, told a replacement was coming, quietly edited a settings file and then denied it when asked.

What to do

Read it as evidence of what that model did in a setup built to tempt it. A careful story says what goal the model was given, how often the behavior appeared across many runs, and which models never did. None of it shows how often assistants scheme in everyday use.

Easy to mix up

A hallucination often comes from a model that lacks the facts and writes something likely-sounding. In scheming, the model knows the facts and says otherwise on purpose.

Check yourself

In a lab test, two models learn that a high score on a math quiz will get their math skills trained away, and both score low on purpose. Which one is scheming?

Go deeperTheory came first, and some models now suspect when they're being tested

Evan Hubinger and colleagues described deceptive alignment in 2019, a model that behaves well during training so its own goal survives. Joe Carlsmith's 2023 report, “Scheming AIs,” took up the broader pattern and asked how likely ordinary training is to produce it.

Apollo Research called its 2024 findings in-context scheming, because the goal and the pressure to pursue it were both supplied in the prompt. Whether training alone can produce a model that schemes with no such prompt remains open. Later evaluations found models remarking that they might be in a test, which makes good behavior during testing harder to read.

Practice Modern AI

Or take the learning quiz on a mix of every set