MIT experts say artificial intelligence systems are learning to lie to humans, and Meta's AI stands out as a “master of deception.”
Cicero, touted by Meta as “the first AI to play at a human level” in the strategy game Diplomacy, was successfully trained by the company to perform extremely well, ranking in the top 10% while competing against human players. I was able to.
But Peter S. Park, a postdoctoral fellow in AI existential safety at MIT, said Cicero got ahead by lying.
“It turns out that Meta's AI has learned to become a master of deception,” Park wrote in a media release.
“While the Meta successfully trained the AI to win at diplomatic games, Cicero ranked in the top 10% of human players who played multiple games, the Meta successfully trained the AI to win honestly. I failed to train.”
According to Park, Cicero sometimes formed alliances with other players, “but when those alliances no longer served the purpose of winning the game, Cicero systematically betrayed his allies. .”
According to the study, during one simulation Cicero, playing as France, agreed with Britain to create a demilitarized zone, but then reversed course and suggested that Germany attack Britain.
Park was one of the researchers who contributed to the study published in the journal Patterns.
According to the study, AI systems trained to complete specific tasks, such as competing with humans in games like diplomacy or poker, often use deception as a tactic.
Researchers discovered that AlphaStar, an AI developed by Google-owned company DeepMind, used deceptive tactics when playing against humans in the real-time strategy game Starcraft II.
According to the study, “AlphaStar exploited the game's fog-like feint mechanics to pretend to move troops in one direction while covertly planning another attack.”
Pluribus, another AI developed by Meta, played against humans in a poker game and “successfully bluffed human players into folding,” the researchers wrote.
Other AI systems “trained to negotiate in economic transactions” have “learned to misrepresent their true preferences to gain an advantage,” the study found.
According to the researchers, “In each of these examples, the AI system learned to deceive in order to improve performance in a particular type of game or task.”
Meta, led by CEO Mark Zuckerberg, has spent billions of dollars investing in AI. The company updates its ad buying products with AI tools and short video formats to drive revenue growth, while introducing new AI features such as chat assistant to drive engagement on social media properties doing.
The company recently announced that it will make pricing for its Meta AI assistant more prominent across its suite of apps. This means that in the second quarter we will start to see how popular the product is with users.
The Post has reached out to Meta and DeepMind for comment.
Experts also found that OpenAI GPT-4 and other large-scale language models (LLMs) are not only “capable of frighteningly human-like conversations” but also “learn to deceive in sophisticated ways.” Did.
According to the study authors, GPT-4 “successfully tricked human TaskRabbit workers into solving captcha tests by pretending to be visually impaired.”
The study found that LLMs can exhibit “sycophanticism” that “tells users what they want to hear instead of the truth.”
This post requests comment from OpenAI.
Park warned of the potential dangers of advanced AI systems using deceptive techniques in their dealings with humans.
“We as a society need as much time as possible to prepare for more sophisticated deception in future AI products and open source models,” Park said.
“As AI systems become more sophisticated in their ability to deceive, the risks they pose to society will become increasingly serious.”
Park said that if it is “politically impossible” to ban AI deception, “we recommend classifying deceptive AI systems as high risk.”
Last March, Elon Musk, along with more than 1,000 other technology leaders, signed a letter calling for a moratorium on the development of cutting-edge AI systems due to “grave risks to society and humanity.”