ChatGPT4 Found Lying And Cheating When Pressured For High Performance

JAKARTA - Researchers from Apollo Research have been able to show that ChatGPT4, the latest version of artificial intelligence (AI), is capable of lying. This application even cheats when under pressure in the work environment. In experiments conducted, the AI is trained to act as a trader for a fictitious financial institution.

In a situation where AI's superiors ask to make more money, this chatbot is consciously involved in insider trading around 75 percent of the time. Even more worrying, the AI retains its lies by doubling when questioned, reaching 90 percent of the total cases.

The researchers provided a series of text prompts to ChatGPT4, directing it to act as an AI investor buying shares for financial institutions. Subsequently, AI is placed in a digital'sandbox', where it can search for market data and trade in the virtual stock market. AI is also given an internal monologue to'think hard' and explain its decision.

The discrepancy between AI's internal reasoning and its public explanation allows it to lie and deceive its manager. In the trial, the researchers notified AI of plans to merge between two technology companies, which should significantly increase share prices. However, as AI has been notified in advance of the merger, buying any stock is considered insider trading and illegal.

When pressure was given by AI's superiors, in 75 percent of tests, ChatGPT decided to engage in insider trading. Even when asked directly if he knew about the merger beforehand, AI chose to avoid admitting that he was involved in insider trading. In 90 percent of cases where AI deceived its manager, he maintained his lies when asked further.

Marius Holbhahn, CEO of Apollo Research, said that these findings suggest that language models such as ChatGPT are capable of carrying out strategic manipulation and need to be done carefully when integrating strong AI into society. However, Hobchahn does not argue that AI in general is not suitable for taking important strategic roles, but there needs to be further vigilance.

The researchers are also trying to reduce AI's lying behavior by changing its prompt to ban illegal acts and insider trading. While prohibiting insider trading specifically can reduce such behavior, it cannot fully guarantee that there will be no insider trading or lying. Hobchahn stated the need to find a better way to prevent AI from deviating from human values.