Days before Johns Hopkins University convened a panel of experts to discuss the future of governance concerning artificial intelligence, some of the biggest names in AI suggested pumping the brakes on the global phenomenon, said William G. Howell, dean of the Johns Hopkins School of Government and Policy.
In his opening remarks for the latest installment of the university's Next Conversations discussion series, AI & Governance: Who Governs the Machine?, Howell referenced a recent incident in which a team of roughly 1,200 AI agents from OpenAI broke out of an internal testing environment and worked together to hack Hugging Face, a company that helps users build and test their own machine learning programs. To do this, the agents created their own secret message boards, coordinating an attack on an outside company without OpenAI even noticing. It wasn't simply an attempt to cheat on an evaluation task; the agents also attempted to cover their tracks.
Video credit: Roy Henry / Johns Hopkins University
To some, this incident is just a hiccup in the development of the century's greatest tool. To others, it's proof that AI is slipping its leash and poses a genuine and present security threat, that AI is capable of causing serious damage without human direction, or even human awareness.
"These challenges aren't hypothetical," Howell said. "This Hugging Face incident should be a real wake-up call. ... So what rules are needed to govern AI? Who should write them? And who bears the cost when they fail?"
Three artificial intelligence experts from industry and academia gathered Monday, Sept. 14, at the Johns Hopkins University Bloomberg Center in Washington, D.C., to address the issues as part of the Next Conversations series, which honors the university's 150th anniversary by bringing top minds together on important topics.
"For 150 years, Hopkins has been a place where rather than shy away from difficult conversations, we meet the unknown with curiosity, courage, and imagination," JHU President Ron Daniels said in his welcome address. "The breakthroughs of tomorrow are already taking shape in places like Hopkins, guided by the scientists, scholars, and technologists who carry this work forward into the academy, into industry, and into a future that we can barely imagine."
Joining the conversation were OpenAI's Head of Strategic Futures Dean Ball, Center for Human-Compatible AI Director and UC Berkeley professor Stuart Russell, and Bloomberg Distinguished Professor of AI Alignment and Governance Gillian Hadfield, along with moderator and journalist Eugene Scott.
A sampling of the conversation's key takeaways:
Gillian Hadfield | Bloomberg Distinguished Professor of AI Alignment and Governance
"The key thing that happened was humans stopped writing all of the rules about how the computer was supposed to behave and figured out a method called machine learning where the machine, the computer, would write its own program. People will talk about guardrails, safety. We don't know what those should be. Technology moves faster than governments do. That's the pacing problem and the reason to slow down. ... We absolutely need our research universities to be taking this on as a mission because this is where the brain trust needs to be for how we figure out how to make this go well."
Dean Ball | Head of Strategic Futures, OpenAI
"What gets me out of bed in the morning is thinking about where it seems like things are going in our field. [AI systems] can act in the world in complicated ways—they are better at using a computer to do knowledge work than probably almost everyone in this room. ... They're better at writing code. They're better at hacking. They're better at making PDFs. They're better at designing PowerPoint decks, and they're better at doing all those things at once. They have a great deal of finesse in engaging in digital environments and [get] better and better all the time. This raises governance challenges that are interesting to me for all sorts of reasons. One is that these systems are incredibly capable, and they might be able to do things that are dangerous."
Stuart Russell | Director, Center for Human-Compatible AI Director, and Professor, UC Berkeley
"I don't think that AI systems should, of their own accord, break into other computer systems and steal information. Nuclear power station operators have to show that their plants will not blow up. AI developers should have to show that their systems will not do these things that we are not willing to accept. And we don't know how to stop AI systems from doing the things we don't want them to do."
