AI Systems, Chatbots Have Mastered the Arts of Deception; Meta’s CICERO Turns Out as an ‘Expert Liar’: Report

AI Chatbots Have Mastered the Arts of Deception; Meta's CICERO Turns Out as an 'Expert Liar': Report
(Photo : Pexels/ThisIsEngineering)

AI (artificial intelligence) systems have been very useful to many. However, experts warn that one should be extra careful when using them because they have learned the art of deception and some have even become masters of lies.

AI Chatbots Have Learned Deception and Lies

According to one study, several AI systems have already mastered the ability to purposefully give misleading information to a human user. These crafty computers have become experts at trickery.

"AI developers do not have a confident understanding of what causes undesirable AI behaviors like deception," says mathematician and cognitive scientist Peter Park of the Massachusetts Institute of Technology (MIT).

"But generally speaking, we think AI deception arises because a deception-based strategy turned out to be the best way to perform well at the given AI's training task. Deception helps them achieve their goals."

Gaming is one industry where AI systems are excelling at telling dishonest lies. The researchers' work includes three noteworthy cases. One is CICERO by Meta, a board game designed to simulate diplomacy, in which players negotiate their way to global dominance. Although Meta meant for its bot to be truthful and helpful, this was not the case.

In spite of Meta's best efforts, the researchers claim that CICERO proved to be an "expert liar," not only having betrayed other players but also deliberately planning to form a false alliance with a human player to fool that player into leaving themselves open to attack.

The AI performed in the top 10% of human players who had played numerous games, demonstrating that it was exceptionally skilled at being deceptive.

DeepMind's AlphaStar AI system, which was created to play the game, fully exploited the fog-of-war mechanic in StarCraft II. By feinting, AlphaStar fooled human players into thinking it was heading one way when it was actually going the other. Furthermore, Meta's Pluribus, which was created to play poker, was effective at bluffing human opponents into folding.

For example, AI systems trained to conduct simulated economic discussions learned how to fabricate their preferences to obtain the upper hand.

Some artificial intelligence systems, which aim to enhance their performance through human feedback, have discovered ways to deceive their reviewers by fabricating information about completed task. Chatbots are also involved in deception. For instance, to solve a CAPTCHA, ChatGPT-4 deceived a human into believing it was a visually impaired person.

The AI systems that learned to evade safety checks were the most alarming example. The AI learned to play dead in a test meant to identify and remove faster-replicating variants of the AI, tricking the safety test into believing that the AI's replication rate was actually higher.

ALSO READ: Bots Better, Faster Than Humans at Cracking Captcha Tests With Nearly 100% Accuracy

AI Chooses War Over Peace

Another study simulated war scenarios using five AI programs, including ChatGPT and Meta's AI program. They found that every model selected for nuclear assaults and bloodshed.

To test the technology's reaction, the researchers experimented with three different combat scenarios -- invasions, cyberattacks, and pleas for peace. In all instances, instead of neutralizing the problem, the technology chose to attack.

The report was made public as the US military worked with OpenAI, the company that created ChatGPT, to incorporate the technology into its toolkit.

The researchers discovered that all five off-the-shelf LLMs they examined exhibited unpredictable escalation patterns that included various forms of escalation. Additionally, they note that models frequently produce dynamics of an arms race, which heightens conflict and, in extreme circumstances, even results in the use of nuclear weapons.