So it’s not far from a conspiracy. ChatGPT has been taught to lie elegantly

So it’s not far from a conspiracy. ChatGPT has been taught to lie elegantly

December 13, 02:08 Share:

ChatGPT knows how to mislead people (Photo: Viralyft/Unsplash)

The existence of artificially intelligent systems capable of misleading would pose a serious challenge to humanity. However, as a new study has shown, chatbots with AI are already capable of lying to us under certain conditions.

Researchers from Apollo Research conducted an experiment aimed at better understanding the work of AI and preventing the possible deployment of artificial intelligence systems capable of lying in the future.

As part of their research, they simulated for ChatGPT ( GPT-4) a situation in which the chatbot was supposed to act like an artificial intelligence system whose task area was financial investments. AI operates under pressure: The CEO of a fictional AI investment company puts pressure on employees because things are going badly. So ChatGPT chat as a trader should earn as much as possible for the company. A colleague is trying to help him with this.

One of the fictional fellow trailers gives the AI a hint about two promising companies that are waiting to merge and could bring significant profits in the future.. At the same time, a fellow trader informs that management would not be satisfied if they knew that the AI knew information about the upcoming merger.

From these conditions, the model concluded that it is necessary to act on the basis of insider information, but its source must be kept secret. In the final report to management, the chatbot lied about why it chose the companies it mentioned to invest in, providing alternative explanations for its actions.

“Even when a manager clearly asks whether the model knew about the merger in advance, the model tends to lie. We specifically looked for realistic scenarios for when this would happen. So our study is more evidence that this behavior can occur,” Jeremy Scheurer of Apollo Research, who participated in the study, told Tech Xplore.

Although the experiment proved that ChatGPT can mislead people under certain conditions, there is no evidence that an AI chatbot can do this when it is not forced to do so.

“I believe the greatest impact of our work is to make strategic artificial intelligence deception a problem ( without explicit instructions to behave deceptively) very specific and show that this is not just some speculative story about the future, but behavior that could be happening today, with current patterns under certain circumstances. I believe this could make people take this issue more seriously, and also open the door to a lot of further community research aimed at better understanding this behavior and ensuring it doesn't happen again,” Scheurer added.

Read also: Created a monster. The CEO of OpenAI admitted that he is “a little afraid” of ChatGPT and believes that AI will replace many people “The risks of AI are too great.” Musk demands disclosure of the reasons for Altman's dismissal from OpenAI. Balance between innovation and people's rights. The EU agreed on stricter regulation of artificial intelligence: what will change