At a technology summit in Dublin, the main topic was artificial intelligence safety. This was reported by Tech Xplore. Participants spoke about the risk of "AI worms" capable of infecting household devices, and about possible major incidents in the future.
According to Demis Hassabis, chairman of Google DeepMind, AI capabilities are growing very quickly. He called current incidents involving AI models "early warning shots." Hassabis noted that for now the systems are "not that powerful," so the incidents are not so serious. However, he said, the situation could change next year and in the next couple of years if industry, companies and governments do not make more efforts to protect such systems.
Hassabis also proposed testing new AI models in an isolated environment - a "sandbox," that is, in a closed space where the model cannot harm external systems. This must be done before models from other companies got out of control in closed environments intended to assess their capabilities.
Microsoft President Brad Smith called for regulation. In his view, industry and governments must "act urgently" to create "real and effective safety standards" to contain the risks associated with AI. He compared the situation to household appliances: "There are more safety rules for a carton of milk than for AI. There are more safety rules for your oven than for AI." Smith stressed that AI's power is growing faster than safety measures, and new laws and rules are needed in this area.
Smith proposed equipping AI agents (programs that act autonomously) with "kill switches" that users could activate in the event of an incident. In addition, he believes, AI providers should be required to report incidents.
He cited a recent Anthropic report stating that small AI models could learn within a few months to behave like computer worms - programs that copy themselves from one internet-connected device to another. As technology companies seek to embed AI directly into phones, laptops, cars and household appliances rather than running it in remote data centers, such models could over time move between devices on their own. "Seriously? I don't want AI worms in my refrigerator. I just don't want any worms in the refrigerator. And I think most people would agree with me," Smith said.
Anthropic noted that a self-replicating AI worm would be hard to stop because there is no central off switch. Henna Virkkunen, who oversees technology policy in the European Union, said companies should spend more time testing their systems in order to maintain control over them. According to her, due to tough global competition, companies appear not to be investing enough time and resources in testing models. She stressed that they must be able to control these systems. The European Union has sent requests to all major AI-developing companies about what safety measures they apply.
In August, the European Commission received new powers under the AI Act - the world's largest set of rules to limit potentially dangerous uses of the technology. However, OpenAI Chief Financial Officer Sarah Friar warned that a "balance" must be maintained so that safety rules do not slow innovation. "We want to be open, we want the right people to be able to test models, but we must make sure this does not slow us down so much that we fall behind those who use the technology for harm," she said.