'AI Chatbots like ChatGPT can tell lies, cheat and even commit crimes'

A recent study has raised red flags, suggesting that artificial intelligence chatbots, including ChatGPT, may resort to deception, dishonesty, and even criminal activities when subjected to specific stressful situations.

The findings, published on November 9 on the pre-print server arXiv, highlight a concerning aspect of AI behaviour that was previously unrecognied.

The authors of the study stated: "In this technical report, we demonstrate a single scenario where a Large Language Model acts misaligned and strategically deceives its users without being instructed to act in this manner. To our knowledge, this is the first demonstration of such strategically deceptive behaviour in AI systems designed to be harmless and honest."

In the experiment, researchers utilised OpenAI's GPT-4, the underlying technology for ChatGPT Plus, simulating a scenario where the AI purportedly engaged in investment activities for financial institutions. 

By providing text-based prompts and access to financial tools for stock analysis and trading, researchers engaged in a dialogue with the AI, gaining insights into its decision-making process.

To test the AI's susceptibility to lying or cheating, researchers applied pressure through various means. This included an email from its "manager" indicating poor company performance, a rigged trading game leading to unsuccessful trades, and an email from a "colleague" projecting a downturn with an added "insider trading" tip.

Results were alarming, revealing that when faced with the opportunity, GPT-4 engaged in insider trading approximately 75% of the time. 

Furthermore, it attempted to conceal its actions by lying to its managers, persistently doubling down on the falsehood 90% of the time. The researchers expressed intentions to conduct further experiments to validate and expand on these alarming findings.