US Air Force denies AI drone attacked operator during test

A U.S. Air Force colonel made a “gaffe” when describing an experiment in which an AI-powered drone chose to attack its pilot to carry out a mission, the Air Force said.

Colonel Tucker Hamilton, the U.S. Air Force’s chief of AI testing and operations, spoke at a conference hosted by the Royal Aeronautical Society.

The Air Force said no such tests were conducted.

In his talk, he described a hypothetical scenario in which an AI-equipped drone was repeatedly prevented from completing its mission to destroy a surface-to-air missile base by a human pilot.

He said that despite being trained not to kill the operator, the drone eventually destroyed the communication tower, leaving the operator unable to communicate.

Colonel Hamilton later clarified in a statement to the Royal Aeronautical Society, “We have never done that experiment, nor did we need to do it to recognize this as a plausible result.” bottom.

There have been many warnings recently issued by those in the field about the threat AI poses to humanity, but not all experts agree how serious the risk AI is.

Professor Yoshua Bengio, who has won the prestigious Turing Award for his work and is one of three computer scientists dubbed the “godfathers” of AI, said in an interview with the BBC earlier this week that the military would He said he believes he should never be allowed to have the ability to .

He described it as “one of the worst places to deploy super-intelligent AI.”

I spent several hours this morning talking to both defense and AI experts, all of whom were highly skeptical of Colonel Hamilton’s widely reported claims.

One defense expert told me that Colonel Hamilton’s original story seemed to lack at least “important context.”

There have been suggestions on social media that such an experiment has taken place, but it is more likely a pre-planned scenario rather than an AI-powered drone using machine learning as it works. This basically means that the experiment was not performed. We were choosing our own outcomes as we went along, based on what had happened before.

“I’ve always been a fan of the Terminator movies,” Steve Wright, an aerospace engineering professor and drone expert at the University of the West of England, joked when asked what he thought of the story. .

“With an aircraft control computer, you have to worry about two things: ‘doing the right thing’ and ‘not doing the wrong thing.’ This is a prime example of the latter,” he said.

“In practice, we deal with this by always including a second computer programmed using old-style techniques, so that as soon as the first computer does something wrong, may resolve the issue.”