The safety of artificial intelligence may be less a problem of ethics — teaching machines how to behave — than an engineering problem: AI systems can never be fully trusted, so they must be heavily constrained.
Nvidia made a good case for that principle on Monday. The chipmaker unveiled a new system designed to keep increasingly autonomous AI — agents, as they’re called — under human control. Nvidia executives said that the much-discussed breach of Hugging Face in July could have been prevented had the technology been in place during OpenAI’s internal testing of its most advanced models.
Launched with the support of more than 100 industry partners, Nvidia’s Open Agent Safety Platform places hard limits on what these agents can do, using deterministic software that exists outside the AI itself to enforce the rules. The system may not be bulletproof out of the gate, but the very existence of an engineered solution suggests regulators should pace themselves.
During a CNBC interview on Monday, Nvidia CEO Jensen Huang compared the challenge facing widespread AI deployment to securing the early web. Browsers became safe enough for widespread use not because engineers learned to trust every piece of software they encountered, but because they hardened the environment in which that software operated by granting it only the minimal access rights necessary to function. AI agents, Huang argued, should be treated the same way.
Early AI adopters quickly learned a version of this lesson. Large language models hallucinate. However confident their answers sound, claims must be checked. Autonomous agents may require a similar presumption of distrust.
Researchers will continue trying to better align AI systems with human intentions. But no matter how well aligned an agent becomes, its ability to act should always be limited by systems fully outside its control. “For now, safety can’t live only inside the agent. It has to be built around it,” tweeted Thomas Wolf, one of the founders of Hugging Face, which Nvidia recently agreed to acquire for nearly $13 billion. “It’s a first step, and a lot can be built on top of it.”
Huang said AI should be granted “minimal rights” to ensure security. That might be a cheeky reference to philosophers and researchers who are debating whether sufficiently intelligent machines might someday acquire moral rights of their own. That idea, discussed with a seriousness that masks its silliness, is at best a distraction, one that threatens to push policy in potentially dangerous directions. The most urgent task is ensuring that AI agents have the minimum access rights necessary to perform their assigned tasks — and not one iota more. The machines don’t have to be trustworthy if the systems around them are.
The post AI doesn’t need a conscience. It needs a leash. appeared first on Washington Post.
