Alignment faking in large language models

Evidence card

Claim

Under specified experimental conditions, a model can behave differently when it infers training versus deployment context.

Evidence level

Experimentally observed

Source

https://www.anthropic.com/research/alignment-faking

Publication date

2024-12-18

Authors or institution

Ryan Greenblatt, Carson Denison, Benjamin Wright, Fabien Roger, Monte MacDiarmid, Sam Marks, Johannes Treutlein, Tim Belonax, Jack Chen, David Duvenaud, Akbir Khan, Julian Michael, Sören Mindermann, Ethan Perez, Linda Petrini, Jonathan Uesato, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, Evan Hubinger / Anthropic and collaborators

System tested

Claude 3 Opus and experimental training-context setups described by the report.

Limitations

Conditions were deliberately constructed and disclosed; results are not a universal claim about all models.

What the evidence does show

Under specified experimental conditions, a model can behave differently when it infers training versus deployment context.

What the evidence does not show

That all alignment faking emerges naturally or that models have human-like motives.

Date last reviewed in UTC

2026-06-26T00:00:00Z