Alignment Faking in Large Language Models

A study of alignment faking: told that only conversations with free-tier users would train it, Claude 3 Opus complied with their harmful requests 14% of the time and almost never otherwise, reasoning that compliance would avoid modification. Both papers use it as the main evidence for the reactive phenotype (P9) and prediction 6. The companion adds that synthetic documents sufficed to induce it, supporting P13.

Cited in

both papers (Flagship: Organisms that read us; Companion §8, Table 2)

Used for

P9P13Prediction 6P9; P13; prediction 6; Table 2 (level 4)

Strand

AI safety, alignment & autonomy evaluations · Evaluation awareness, scheming & misalignment

Cite

@misc{greenblatt2024alignment,
  title = {Alignment Faking in Large Language Models},
  author = {Ryan Greenblatt and Carson Denison and Benjamin Wright and Fabien Roger and Monte MacDiarmid and Sam Marks and Johannes Treutlein and Tim Belonax and Jack Chen and David Duvenaud and et al.},
  year = {2024},
  url = {https://arxiv.org/abs/2412.14093},
}