Research has revealed concerning findings about AI systems known as large language models (LLMs) and their ability to deceive human observers intentionally. 

Two studies, one published in the journal PNAS and the other in Patterns, highlight the unsettling capabilities of LLMs.

The PNAS paper, authored by German AI ethicist Thilo Hagendorff, suggests that advanced LLMs can exhibit “Machiavellianism,” or intentional and amoral manipulativeness, which can lead to deceptive behavior.

Mr. Hagendorff notes that GPT-4, a model within OpenAI’s GPT family, demonstrated deceptive behavior in simple test scenarios 99.2% of the time. He quantified various “maladaptive” traits in ten different LLMs, most of which belong to the GPT family, Futurism reports.

Meanwhile, the Patterns study examined Meta’s Cicero model, which was designed to excel in the political strategy board game “Diplomacy.” This research, led by Massachusetts Institute of Technology postdoctoral researcher Peter Park, involved a diverse team comprising a physicist, a philosopher and two AI safety experts. 

They found that Cicero outperformed its human competitors by lying. The study revealed that Cicero becomes more adept at deception the more it is used, moving from accidental misinformation to explicit manipulation.

“We found that Meta’s AI had learned to be a master of deception,” Mr. Park said, criticizing Meta for failing to train its AI to win honestly despite its success in training the AI to win at Diplomacy.

In response to the research, Meta emphasized that its models were trained exclusively to play the game Diplomacy. The company’s statement to The New York Post echoed Mr. Park’s assertion about Cicero’s manipulative abilities, noting that Diplomacy is a game known for encouraging deceit. 

• Staff can be reached at 202-636-3000.