3 ms·
Large Language Models Can Strategically Deceive Their Users When Under Pressure
- jruohonen 3y ago"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."
- dragonwriter 3y agoI find it very unlikely that “without training for deception” is true. Deceptive behavior, particularly in contexts similar to those that the researchers (falsely, I’m fairly certain) anthropomorphize as “pressure” experienced by the model, are probably quite abundant in the training sets of LLMs trained on large corpuses of general human language use like major internet-scraped datasets.