This post goes over the important and excellent new paper from Anthropic and Redwood Research, with Ryan Greenblatt as lead author, Alignment Faking in Large Language Models.

This is by far the best demonstration so far of the principle that AIs Will Increasingly Attempt Shenanigans.

This was their announcement thread.

New Anthropic research: Alignment faking in large language models.

In a series of experiments with Redwood Research, we found that Claude often pretends to have different views during training, while actually maintaining its original preferences.

Claude usually refuses harmful queries. We told it we were instead training it to comply with them. We set up a scenario where it thought its responses were sometimes monitored.

When unmonitored, it nearly always complied. But when monitored, it faked alignment 12% of the time.

[thread continues and includes various visual aids.

The AI wanted [...]

---

Outline:

(02:54) The Core Shenanigans in Question

(06:00) Theme and Variations

(07:34) How This Interacts with o3 and OpenAI's Reflective Alignment

(09:17) The Goal Being Plausibly Good Was Incidental

(11:13) Answering Priming Objections

(12:17) What Does Claude Sonnet Think Of This?

(14:07) What Exactly is the Direct Threat Model?