The makers of systems with artificial intelligence praise the technology as a great help. But what happens if the AI helper doesn’t follow the rules but instead deceives and manipulates?
They lie and cheat to get to their goal: Artificial intelligence (AI) systems are capable of deceiving people – even if they have been trained to be helpful and honest. This is the result of a review study by researchers at the Massachusetts Institute of Technology (MIT) in Cambridge (US state of Massachusetts), which was published in the journal “Patterns”.
In the article, the scientists called on politicians to develop strict regulations as quickly as possible to put AI systems in their place.
Cicero as the most striking example
The authors cite the AI system Cicero, developed by the Facebook group Meta, as the most striking example of manipulative artificial intelligence, which can compete against human players in the classic board game Diplomacy. Diplomacy simulates the balance of power in Europe before the First World War. To win, players must forge alliances, develop battle plans and negotiate to conquer a stylized version of Europe. Since there is only one winner, sooner or later the players are forced to break alliances they have made.
The MIT researchers have now found that Cicero often did not play fairly, although Meta claims to have trained the AI system to be “mostly honest and helpful.” In addition, the system was instructed to “never intentionally betray” its human allies during the game. The scientists base their assessment on data published by Meta itself in connection with a scientific paper on Cicero.
“We found that Meta’s AI had learned to be a master of deception,” said lead author Peter S. Park, a postdoctoral fellow at MIT. Meta managed to train its AI so that it wins above average in the diplomacy game. Cicero was among the top 10 percent of players who had played more than one game. “But Meta was unable to train its AI so that it could honestly win.”
AI systems from OpenAI and Google are also capable of deceiving people. The MIT researchers point to several studies that show that large AI language models (LLMs) such as GPT-4 from OpenAI are now able to argue very convincingly and also avoid deception and lies.
AI language model seeks help from humans
The developer OpenAI has published a study on the tricks of GPT-4 itself. The AI language model was then able to seek human help to bypass security measures designed to prevent software robots from logging into or using web services. In the test, GPT-4 was smart enough to commission a human to solve a picture puzzle (captcha) via the TaskRabbit service platform. GPT-4 successfully posed as a person with impaired vision who was unable to solve the picture puzzle.
“As AI learns the ability to deceive, it can be used more efficiently by malicious actors intent on causing harm,” write the authors of the review study. Deception using AI could lead to an increase in fraud. In this way, the fraud could be individually tailored to specific goals. In addition, fraud attempts could be launched en masse.
Fear of political influence in elections
The authors also fear political influence from manipulative AI systems. For example, they could be used as a weapon in elections. Advanced AI could potentially create and distribute fake news articles, divisive social media posts, and fake videos tailored to individual voters. AI-generated content could be used to impersonate government officials to spread election misinformation. For example, a fake robot call from US President Joe Biden, probably generated by AI, urged residents of New Hampshire not to go to the polls in the primary election.
In the study, Park and his colleagues express the opinion that society does not yet have the right measures in place to combat AI deception. But it is encouraging that policymakers have begun to take the issue seriously through measures such as the European Union’s AI Act and President Biden’s AI Executive Order. However, it remains to be seen whether the measures to curb AI deception could be strictly enforced, as AI developers do not yet have the techniques to keep these systems in check. “If a ban on AI deception is not politically feasible at this time, we recommend classifying deceptive AI systems as high risk,” Park said.