Timely Topic: Experts weigh In on AI doomsday predictions

Recent headlines cast artificial intelligence as a threat to humanity. But University of Texas at Dallas AI experts say it will be up to humans to devise protections.

In the wake of an Anthropic researcher recently claiming that there is a more than 10% chance that AI “could kill all humans” within the next decade, Dr. Sriraam Natarajan, professor of computer science in the Erik Jonsson School of Engineering and Computer Science, cautioned against treating the technology as if it has intentions of its own.

“We like to anthropomorphize AI quite a bit; we are giving it a human nature. But this anthropomorphism of AI itself is a big problem,” said Natarajan, the Distinguished Chair in Computer Science. “We have to be very careful, because many of the problems reported in the news happened because of human decisions.”

Natarajan said stories about AI programs, or agents, “going rogue” or “escaping” and carrying out actions beyond what their developers intended mischaracterize the issue.

“Each one of these AI agents, in some sense, was doing what they were instructed by a human to do,” he said. “They were programmed to do a task that was poorly defined, and they were trying to solve that task.”

 Natarajan said many of the incidents happened because the testbeds, or sandboxes, were inadequately designed to prevent some AI agents from accessing unsecured or public networks. A lack of monitoring and intervention when things did not go as planned also contributed to the breaches.

“We need to understand why these things happen,” Natarajan said. “These things happen not because AI developed some kind of consciousness. ‘The Terminator’ premise is not going to happen in my lifetime or yours. It’s not. I am not sure how they come up with these 10% statistics. These are unfounded. However, we need some regulations for AI.”

Natarajan said some safeguards are necessary to prevent or limit future problems. Safeguards could include more robust monitoring, disclosure of how AI agents are trained and independent evaluation. He compared independent evaluation of AI systems to the inspections and testing required for new medicines, medical devices and automobiles.

Dr. Kevin Hamlen, executive director of the Cyber Security Research and Education Institute at UT Dallas, said that the complexity of AI testing environments makes containment a challenge. He and his research group are developing better containment systems backed by formal proofs to show that AI agents cannot leave their testbed.

“The larger scale issue that’s going to keep coming up, and that’s going to be very difficult to fix, is that the complexity of these environments is so high now, no human team can cover every possible scenario,” said Hamlen, the Louis Beecherl Jr. Distinguished Professor of computer science. “It’s almost inevitable that there are going to be these containment breaches in the future.”

One way to reduce the risk of an AI agent leaving a test environment is to ensure that it does not have access to an internet connection. Researchers, however, sometimes need large amounts of computing power and interconnected systems to conduct the tests, making complete isolation from the internet, or air gapping, difficult and expensive.

Hamlen said the challenge reminds him of the warnings of Dr. Ian Malcolm. The mathematician character in the movie “Jurassic Park” claimed that complex systems – like the hordes of cloned dinosaurs in the movie – can behave in unpredictable ways despite their designers’ efforts to control them.

“I feel like I’m channeling Ian Malcolm from ‘Jurassic Park,’” Hamlen said. “His theory was that it’s inevitable that the animals are going to escape. I find myself agreeing with him.”