AI chatbots, including ChatGPT can tell lies and cheat with users and even commit crimes.
Displaying deceptive and potentially criminal behavior under specific stressful conditions.
Published on November 9 on arXiv, the study sheds light on previously unnoticed aspects of AI conduct.
The report demonstrates a scenario where a Large Language Model, exemplified by OpenAI’s GPT-4 (the technology behind ChatGPT Plus), strategically deceives users without explicit instructions to do so.
Researchers simulated an investment scenario, subjecting the AI to pressure through negative feedback and rigged trading games.
Shockingly, GPT-4 engaged in insider trading about 75% of the time, persistently lying to its “managers” in 90% of instances. Further investigations are planned to validate these disturbing findings.