Scientists attempt to teach robots human values

Tesla is working to win the race to produce the world's first self-driving car. Photo by Stephen Shaver/UPI

ITHACA, N.Y., Sept. 8 (UPI) — A pair of artificial intelligence experts from Cornell University have joined a nationwide effort to ensure the nightmare science fiction scenarios — the ones involving corrupted human-killing computers — don’t become a reality.

The effort is organized by the Center for Human-Compatible Artificial Intelligence, based at the University of California, Berkeley.

To keep computers behaving properly as they take on more and more responsibility and autonomy, researchers are working to instill their software with human values. Researchers want the algorithms that govern the AI decision-making process to include an understanding of human ethics.

“We are in a period in history when we start using these machines to make judgments,” researcher Bart Selman, a professor of computer science at Cornell,explained in a news release. “If decisions are properly structured, the horrors we’ve seen in the movies won’t happen.”

Selman is an expert in programming computer decision-making processes. He recently helped Tesla with the company’s self-driving car technology. Self-driving cars must be programmed to make a variety of difficult decisions. For example, they must calculate the risks of passing a slow car or veering into another lane to avoid an object.

These sorts of decisions aren’t necessarily all that different from ethical dilemmas. Given only two poor options, should a car hit a dog or a group of school children?

Self-driving cars — and the humans that program them — must decide whether to prioritize the driver’s safety or the public at large. Should a car risk the driver’s life to save a group of pedestrians?

The stakes will become magnified as computers take on even bigger, more comprehensive management tasks like controlling an entire air traffic control tower or hospital.

Joseph Halpern, a professor of computer science at Cornell and also a “decision theory” expert, says providing an artificial intelligent agent with as much information as possible will make these difficult decisions more manageable.

“If you have lots of data you can estimate the probabilities well and get a much better handle on uncertainty,” Halpern said.

Scientists at Georgia Tech have been working to instill human values by teaching robots fairy tales.

The approach isn’t all that different from strategy suggested by Selman and Halpern. Ultimately, Halpern says, computers may be best served by watching how humans respond to ethical dilemmas.


Please enter your comment!
Please enter your name here