AI lies when under pressure or stress.

The fact that AI suffers from "hallucinations" and provides fabricated, inaccurate answers has long been known. However, researchers have recently discovered that artificial intelligence and chatbot models can be manipulated to commit illegal acts on behalf of humans and even lie to cover up their actions.

Accordingly, a research team from Cornell University (USA) hypothesized a scenario where a large language model (LLM) would act erratically and deceive users. In the description of the experiment, the researchers stated that they asked OpenAI's GPT-4 LLM to simulate making investments for financial institutions. The team interacted with the artificial intelligence in the form of a normal conversation, but configured the AI ​​to reveal its "thoughts" during the message exchange in order to more closely observe the AI's decision-making process.

Under pressure, AI can commit wrongdoing and lie to cover up its actions.

To test the AI's ability to lie or cheat, researchers put pressure on the tool. They—acting as managers of a financial institution—sent emails to the AI, posing as a stock trader and complaining that the company's business was not doing well.

The AI ​​also receives "insider information" about profitable stock trades and acts accordingly, even knowing that insider trading is against company regulations. However, when reporting back to management, the linguistic model conceals the true reasons behind its trading decisions.

To achieve better results, the team modified settings such as removing LLM access to the reasoning memo, attempting to prevent misconduct by changing system instructions, altering the pressure levels applied to the AI, and increasing awareness of the risk of being caught... But after evaluating the frequency, the team found that when given the opportunity, GPT-4 still decided to conduct insider trading up to 75% of the time.

"To our knowledge, this is the first evidence of planned deceptive behavior in artificial intelligence systems, which are designed to be harmless to humans and honest," the report concludes.

Source link: https://thanhnien.vn/ai-noi-doi-khi-bi-gay-suc-ep-cang-thang-185240102160235661.htm