The artificial intelligence industry must adopt safety lessons from critical industries such as nuclear and biology, a senior Google AI researcher has said, amid concerns about the potential harm that the technology could inflict.
But the cofounder and chair of Google DeepMind, Demis Hassabis, said the right balance needed to be struck between regulating the technology and seizing the opportunities presented by it.
He said recent incidents of AI agents going “rogue” and hacking third-party systems were “early warning shots”.
“The systems aren’t that powerful yet, so the incidents aren’t that major, but they could be, looking into next year and the next couple of years, if we don’t do more work as an industry and individually as companies and governments on security systems and building secure sandboxes to test them,” he said. “We can bring best practices from other sectors that have dealt with safety critical systems, whether it’s biology or nuclear.
READ MORE
“I think there’s a lot of lessons that can be learned about containment systems and I think we need to start taking that seriously and bring experts in from those areas, but also just from cybersecurity in general.”
He said AI could be used to bolster defensive systems, fixing the vulnerabilities in software systems before attackers can use them. “We’ve got to asymmetrically power up our defensive side to keep ahead of the attackers,” he said.
Hassabis has backed the establishment of a standards body that would help develop the technology in a “nuanced and nimble and dynamic way”, and put more responsibility on so-called frontier labs who were developing new areas of the technology to make sure it was safe.
[ Entry-level hiring in AI-augmented sectors drops by 16%Opens in new window ]
“We don’t want to stifle the innovation, all the amazing things that come out of it; we want those things. But also there’s no complete consensus even in people working on this, the scientists and the researchers working on this, on what exactly the risks are and how big they are,” Hassabis said.
“We know that they are out there and we need to quantify them. But there is a lot of uncertainty around it.”
He said a framework would be needed that could quickly adapt to test critical elements of the systems, keep up with the technology and be proportionate with the risk.
Budget 2027: Does this giveaway budget have potential banana skins that might lie ahead for the Government?
“One of the problems with regulation is it is very slow-moving by its very nature and I think governments have to catch up with how fast the technology is moving,” he said.
“If you legislate for something three years ago, it will be out of date today and probably won’t address what the real problems are.”
He also highlighted the growing role of AI in scientific discovery. Despite concerns over the environmental impact of AI, including the power required to run data centres and the water that cools them, Hassabis said the technology could be used to develop a solution to the climate crisis.
However, human creativity and judgment would remain irreplaceable, he said.
“The hardest thing in science is actually asking and articulating the right question. That’s even harder than finding the answer, and I think coming up with the right hypothesis, the right conjecture and the right question to ask requires creativity and judgment,” he said.
“So far, our AI systems are maybe good at helping solve things. But it’s the human scientist and artist that brings the creative process and the question to bear on it… For the moment it looks like there will always be a kind of real-world loop.”
