Computer scientists who helped build the foundations of todays artificial intelligence technology are warning of its dangers, but that doesnt mean they agree on what those dangers are or how to prevent them.
Humanitys survival is threatened when smart things can outsmart us, so-called Godfather of AI Geoffrey Hinton said at a conference Wednesday at the Massachusetts Institute of Technology.
After retiring from Google so he could speak more freely, the 75-year-old Hinton said hes recently changed his views about the reasoning capabilities of the computer systems hes spent a lifetime researching.
These things will have learned from us, by reading all the novels that ever were and everything Machiavelli ever wrote, how to manipulate people, Hinton said, addressing the crowd attending MIT Technology Reviews EmTech Digital conference from his home via video. Even if they cant directly pull levers, they can certainly get us to pull levers.
I wish I had a nice simple solution for this, but I dont, he added. Im not sure there is a solution.
Fellow AI pioneer Yoshua Bengio, co-winner with Hinton of the top computer science prize, told The Associated Press on Wednesday that hes pretty much aligned with Hintons concerns brought on by chatbots such as ChatGPT and related technology, but worries that to simply say Were doomed is not going to help.
The main difference, I would say, is hes kind of a pessimistic person, and Im more on the optimistic side, said Bengio, a professor at the University of Montreal. I do think that the dangers the short-term ones, the long-term ones are very serious and need to be taken seriously by not just a few researchers but governments and the population.
There are plenty of signs that governments are listening. The White House has called in the CEOs of Google, Microsoft and ChatGPT-maker OpenAI to meet Thursday with Vice President Kamala Harris in whats being described by officials as a frank discussion on how to mitigate both the near-term and long-term risks of their technology. European lawmakers are also accelerating negotiations to pass sweeping new AI rules.
But all the talk of the most dire future dangers has some worried that hype around superhuman machines which dont yet exist is distracting from attempts to set practical safeguards on current AI products that are largely unregulated.
Margaret Mitchell, a former leader on Googles AI ethics team, said shes upset that Hinton didnt speak out during his decade in a position of power at Google, especially after the 2020 ouster of prominent Black scientist Timnit Gebru, who had studied the harms of large language models before they were widely commercialized into products such as ChatGPT and Googles Bard.
Its a privilege that he gets to jump from the realities of the propagation of discrimination now, the propagation of hate language, the toxicity and nonconsensual pornography of women, all of these issues that are actively harming people who are marginalized in tech, said Mitchell, who was also forced out of Google in the aftermath of Gebrus departure. Hes skipping over all of those things to worry about something farther off.
Bengio, Hinton and a third researcher, Yann LeCun, who works at Facebook parent Meta, were all awarded the Turing Prize in 2019 for their breakthroughs in the field of artificial neural networks, instrumental to the development of todays AI applications such as ChatGPT.
Bengio, the only one of the three who didnt take a job with a tech giant, has voiced concerns for years about near-term AI risks, including job market destabilization, automated weaponry and the dangers of biased data sets.
But those concerns have grown recently, leading Bengio to join other computer scientists and tech business leaders like Elon Musk and Apple co-founder Steve Wozniak in calling for a six-month pause on developing AI systems more powerful than OpenAIs latest model, GPT-4.
Bengio said Wednesday he believes the latest AI language models already pass the Turing test named after British codebreaker and AI pioneer Alan Turings method introduced in 1950 to measure when AI becomes indistinguishable from a human at least on the surface.
Thats a milestone that can have drastic consequences if were not careful, Bengio said. My main concern is how they can be exploited for nefarious purposes to destabilize democracies, for cyber attacks, disinformation. You can have a conversation with these systems and think that youre interacting with a human. Theyre difficult to spot.
Where researchers are less likely to agree is on how current AI language systems which have many limitations, including a tendency to fabricate information will actually get smarter than humans.
Aidan Gomez was one of the co-authors of the pioneering 2017 paper that introduced a so-called transformer technique the T at the end of ChatGPT for improving the performance of machine-learning systems, especially in how they learn from passages of text. Then just a 20-year-old intern at Google, Gomez remembers laying on a couch at the companys California headquarters when his team sent out the paper around 3 a.m. when it was due.
Aidan, this is going to be so huge, he remembers a colleague telling him, of the work thats since helped lead to new systems that can generate humanlike prose and imagery.
Six years later and now CEO of his own AI company called Cohere, which Hinton has invested in, Gomez is enthused about the potential applications of these systems but bothered by fearmongering he says is detached from the reality of their true capabilities and relies on extraordinary leaps of imagination and reasoning.
The notion that these models are somehow gonna get access to our nuclear weapons and launch some sort of extinction-level event is not a productive discourse to have, Gomez said. Its harmful to those real pragmatic policy efforts that are trying to do something good.
