How do we know AI will obey? Maybe look at how they argue with themselves

Experts believe that seeing artificial intelligence systems (AI) argue with each other will help us understand their intentions.

Someday, it may be perfectly normal for an AI system to fight itself.

The idea comes from researchers at a non-profit organization founded by some Silicon Valley leaders, including Y Combinator Sam Altman's partner, Reid Hoffman's LinkedIn chairman, Facebook board member. and is the founder of Palantir Peter Thiel, and Elon Musk, CEO of Tesla and SpaceX.

Earlier, OpenAI researchers have shown that these can sometimes develop unexpected and unwanted behaviors. For example, in a computer game, an AI device can find a way to "hack" to get a higher score. In some cases, people can monitor the training process. But if the AI ​​program is doing something complicated, monitoring is not feasible. So the researchers suggest that instead of human supervision, let the two AI systems discuss, fight with each other on an issue.

Self-learning AI systems can sometimes develop unexpected and unwanted behaviors.

"We believe this approach or something like this can help us train the AI ​​system to perform more cognitive tasks than human ability, while still matching the preferences of people, " researchers wrote in a blog post.

For example, an AI system is designed to protect against human or AI hackers. To prevent the system from doing anything harmful or unethical, it may be necessary to challenge the AI ​​system to explain logic for a specific action. That logic may be too complicated for people to understand, so the researchers suggest that another AI should argue about the logic, behavior of the first AI system, and that people will observe.

Allowing AI programs to argue with each other requires more complex technology than currently. So, so far, OpenAI researchers only explore ideas with a few extremely simple examples. For example, let two AI systems try to convince an observer about a mysterious character, by slowly revealing the individual elements.

Researchers have created a website, in which anyone, with two people, can try to play role-playing in the controversial AI system, while a third person plays an observation and review role. judge. The two will compete together to convince the judge about the nature of an image while highlighting its parts. And finally, the observer (judge) can easily identify who is honest.

Vincent Conitzer, a researcher at Duke University, studies ethical issues related to AI, saying the project is at an early stage but very promising."Creating AI systems that can explain their decisions is a challenging research program ," he said. "If successful, it can contribute greatly to the responsible use of AI."

Despite some strange statements by Elon Musk (Elon Musk is an OpenAI sponsor and a member of the board of directors), we still have a long way to go to have an AI system possible. the ability to deceive and deceive us as described in the movies Ex Machina and Her.

There is still a long way to have an AI system capable of deceiving and deceiving us.

According to Technology Review, however, some AI researchers are looking for ways to ensure AI technology will not work unexpectedly and unpredictably. This will be very important when AI programs become more complex and confusing.

Ariel Procaccia, CMU's computer science professor, who studies how decision-making systems are autonomous, said: "I think the idea of ​​value association is through a very interesting debate. and have useful ".

However, Procaccia notes that everything is still very preliminary, and that idea may even contain a fundamental contradiction."In order to debate valuable questions in an easy-to-understand way for a human judge, AI systems may have to understand human values ​​from the beginning , " he said.

Iyad Rawan, a researcher at MIT's communications lab, added that researchers need to be careful."I think some difficulties will arise very quickly , " he said. "The first is how the debate automatically takes place in natural language, it's still an unsolved problem."