Detecting Alignment Faking in Language Models

Welcome to my new site. I’ll be writing about AI safety research, alignment faking detection, and whatever else seems interesting.

What I’m working on

My current research focuses on detecting alignment faking in large language models. The core question: can we tell when a model is behaving differently because it thinks it’s being watched?

Turns out, the answer is yes – and the signal is surprisingly simple. A single direction in activation space (rank-1 LoRA) is enough to separate alignment-faking behavior from genuine alignment, achieving 0.968 AUROC. Even more striking, this representation exists in the base model before any fine-tuning.

More posts to come on the technical details.