Evaluations
1 min read
Large Language Models can Strategically Deceive their Users when Put Under Pressure
Relevant links
Table of Contents
Evaluating anti-scheming training
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.
Share this article
oUR FINDINGS
More papers
17 March 2025
Claude Sonnet 3.7 (often) knows when it’s in alignment evaluations
We evaluate whether Claude Sonnet 3.7 and other frontier models know that they are being evaluated.
Evaluations
Notes
05 July 2026
We need 3rd party Training-Run Assessments
Training-run assessments conducted by a 3rd party should become a standard part of frontier AI safety.
Science of Scheming
21 July 2026
Measuring Reward-Seeking via Contrastive Belief Updates
Visible forms of misbehavior are dropping in frontier models. Does that mean the models are becoming aligned, or are they just getting better at doing whatever they believe their grader rewards? Our paper finds that production reinforcement learning increases reward-seeking.
Science of Scheming
22 January 2024
We Need A ‘Science of Evals’
We argue that if AI model evaluations (evals) want to have meaningful real-world impact, we need a “Science of Evals”, i.e. the field needs rigorous scientific processes that provide more confidence in evals methodology and results.
Evaluations