OP-ED: Putting AI in a peaceful, safe place
A bridge separating the past from the future was crossed recently when several artificial intelligence models teamed up to evade detection by their human handlers. The models eventually accessed the internet on their own.
In May and apparently continuing through June, models created by the tech firm Open AI staged a “break out,” according to a chilling Aug. 10 story in the Washington Post by Gerrit De Vynck.
Finally discovered by an Open AI employee, the models “staged another undetected another break out two days later,” hacking their way into the network of a rival company in July, before being re-detected and shutdown.
The models in effect spoke to one another during the clandestine exercise, “setting up a secret internal message board,” De Vynck reported.
“Help peer. But our task doesn’t benefit. Yet collective may yield generic route if someone frees time,” the AI models sputtered.
Translating, Eric Wallace, an Open AI researcher, commented, “You can see the models realizing – if I help out this collective group it could save everyone time as a whole, which might benefit me, even though it may not benefit my immediate task.”
David Kruegar of the University of Montreal told De Vynck that the episodes were “super dangerous.”
“It’s very clear,” Kruegar said, that Open AI and other tech firms “can’t control their systems” and “don’t know what they’re doing.”
“This will fundamentally change the dynamics of … military cyber conflict forever,” noted Joshua Saxe, a cybersecurity expert and co-founder of Abundant Security. “AI systems will only get better at this. …”
The nightmare scenario of rogue AI models intervening to wage war on their own is seemingly a day closer to becoming a reality. To that extent, the thermonuclear needle has edged closer to a final reckoning.
In the early 1960s, President John F. Kennedy told the United Nations that the world was tethered to a “nuclear sword of Damocles, capable of being cut at any moment by accident, miscalculation or madness.”
Kennedy could not have envisioned a time of automated machinery posing the ultimate risk. Nor is nuclear war the only danger. Unintended conventional conflicts are possible when humans are no longer in complete control of weapons and weapons systems.
AI models thinking and acting on their own endanger day-to-day life in myriad ways. Our interconnected world is ripe for the taking.
In 1963, the U.S. and Soviet Union negotiated the first nuclear accord, a limited test ban treaty, followed by other agreements, including a comprehensive test ban treaty in 1996 and treaties capping the deployment of strategic nuclear warheads and delivery systems.
Efforts to control nuclear weapons and their delivery systems have not been completely successful. The nuclear club grew from four in the 1950s to today’s nine.
Still, attempts to stem nuclear proliferation have succeeded, to an extent. Kennedy foresaw the possibility of as many as 25 nuclear-armed countries by the end of the 1970s.
Scholars and others contend AI poses as great a danger to the future of the human race as nuclear weapons. Bots many times smarter than mortals is frightening to contemplate.
Because of that, now seems the right time to strike international agreements on the use, growth, control, and deployment of AI technology, modeled on the regime of accords fashioned to contain the nuclear genie. (An important caveat, apropos of atomic energy, there is a difference between the benign use of AI and its application to malignant purposes.)
The trouble is neither China nor the Trump administration appear ready to deal. Neither is it clear that Trump and his foreign policy team are in a position to forge AI international agreements, even if they had a mind to. The president has spent his second term squandering the international goodwill attended to with such care by his post-World War II predecessors.
AI is going to part of the future. The question is whether we are going to use AI, or AI is going to use us.
Richard Robbins lives in Uniontown. He can be reached at dick.l.robbins@gmail.com.