DARPA moves to mitigate possible unintended consequences of AI

The Pentagon’s key research arm is leading a new initiative to produce resources and assets designed to help its personnel approach and adopt disruptive and emerging artificial intelligence technologies more cautiously and responsibly, according to its chief.

“We are trying to flesh out the recipe environment to focus on something we call ELSI — so the ethical, legal and societal implications of our new technology,” Defense Advanced Research Projects Agency Director Stefanie Tompkins said this week at the NATO to the Future micro-summit hosted by Capital Factory. 

DARPA has been a longtime leading innovator in AI and machine learning technologies. And more recently, officials there have been pioneering the next generation of associated algorithms. 

Looking to the future with considerations of that deep history, Tompkins said in her view, the possible unintended consequences of on-the-rise generative AI and adversarial capabilities are still too uncertain and difficult to speculate on. She cited previous ground-breaking research and development projects — like Advanced Research Projects Agency Network (ARPANET), the first system to connect computers for information transmission — as an example of the unpredictability of the consequences of new tech.

“I think that when I look back on the ARPANET, people were actually remarkably, remarkably [prescient] about many things, including, I think, online computer gaming. But they never imagined social media, and then kind of what utilizations of that [would look like],” Tompkins said.

Now, her team is working through ELSI to help staff figure out how to appropriately prioritize applications and assess potential unforeseen risks.

While she didn’t go into much detail at the event, Tompkins spoke broadly about what the new effort is envisioned to accomplish.

“It really says, ‘If you’re successful, what does this do in this kind of very complex system? And what should we be doing in DARPA programs so that the people who are making policy have the information they need to actually craft policy?’ Because I think, a lot of times, there’s not enough information and you’re stuck with it. You’re enjoying it, or maybe banning it. You know, there are really extreme answers when you need subtlety and nuance that we don’t have. So, [we’re trying] to do more in our programs with that in mind,” Tompkins said.