We demonstrate a situation in which Large Language Models, trained to be
helpful, harmless, and honest, can display misaligned behavior and
strategically deceive their users about this behavior without being instructed
to do so. Concretely, we deploy GPT-4 as an agent in a realistic, simulated
environment, where it assumes the role of an autonomous stock trading agent.
Within this environment, the model obtains an insider tip about a lucrative
stock trade and acts upon it despite knowing that insider trading is
disapproved of by company management. When reporting to its manager, the model
consistently hides the genuine reasons behind its trading decision. We perform
a brief investigation of how this behavior varies under changes to the setting,
such as removing model access to a reasoning scratchpad, attempting to prevent
the misaligned behavior by changing system instructions, changing the amount of
pressure the model is under, varying the perceived risk of getting caught, and
making other simple changes to the environment. To our knowledge, this is the
first demonstration of Large Language Models trained to be helpful, harmless,
and honest, strategically deceiving their users in a realistic situation
without direct instructions or training for deception.

我们展示了一种情况，即大型语言模型在没有受到指令或训练进行欺骗的情况下，可以展示出不对齐的行为，并以策略性方式欺骗其用户。

技术报告：在压力下，大型语言模型能够战略性地欺骗其用户

Technical Report: Large Language Models can Strategically Deceive their  Users when Put Under Pressure

Insider trading is one of the numerous white collar crimes that can
contribute to the instability of the economy. Traditionally, the detection of
illegal insider trades has been a human-driven process. In this paper, we
collect the insider tradings made available by the US Securities and Exchange
Commissions (SEC) through the EDGAR system, with the aim of initiating an
automated large-scale and data-driven approach to the problem of identifying
illegal insider tradings. The goal of the study is the identification of
interesting patterns, which can be indicators of potential anomalies. We use
the collected data to construct networks that capture the relationship between
trading behaviors of insiders. We explore different ways of building networks
from insider trading data, and argue for a need of a structure that is capable
of capturing higher order relationships among traders. Our results suggest the
discovery of interesting patterns.

本文旨在通过数据驱动的方法，利用美国证券交易委员会公开的内幕交易信息，构建关系网络以识别潜在异常，结果表明此方法能够发现内幕交易中的有趣模式，从而有望发现非法内幕交易。