TLDR

Models often behave dishonestly without acquiring a coherent deceptive disposition.
We trained some mid-sized models on their own plausible but false reasoning.
True and false training usually produced nearly identical downstream effects.
Even statements contradicting latent knowledge transferred only weakly to unrelated dishonesty.
General deception may require agency, persistent private information, and successful concealment over time.

Introduction

Current frontier models are mundanely misaligned. That is, they oversell work, claim completion too early, reward hack in ways users would reasonably call dishonest. But they do not yet seem dishonest in the way a person is dishonest. Models have this sort of sheepishness, acting abashed when called out, and then, as if, forgetting, doing it again. Humans caught hacking and overselling would typically go further and dissemble, or be defensive.

What are the generalization boundaries of dishonesty? Overselling your work is kind of like misrepresenting it, reward hacking is kind of like covering it up. And we know language models love to generalize. Yet models don't seem to make the jump from "behaves in ways that look dishonest" to "is dishonest" in the way a person would be, lacking a sort of coherent motivation. If we are correct that [...]