Opinion | The Luddite and the engineer

El-Sayed’s smorgasbord of fanciful proposals embraces statism. He wants to “freeze the training and release” of models until a new Cabinet-level agency can impose a novel regulatory regime. He supports public ownership of the frontier labs. And he naively envisions a treaty with China “to pause and ban superintelligence.” El-Sayed would also require kill switches to be built into all AI systems and impose “severe criminal penalties,” including prison time, for AI executives who break his rules.

Public anxiety is palpable, and it’s fair for politicians to take voters’ concerns seriously. But El-Sayed’s approach, which echoes the policies being peddled by Sen. Bernie Sanders (I-Vermont), preys on people’s fears. He wouldn’t erect guardrails so much as throw sand in the gears of human progress.

For inspiration, he might instead look to a recent short essay by Microsoft’s chief executive, Satya Nadella. Nadella is no Pollyanna about AI’s dangers. He acknowledges that we do not fully understand how these machines arrive at their decisions and that the risks of letting them run unchecked are grave. But instead of demanding a halt to their development, he approaches the problem as an engineer. “We need to separate the supply of intelligence from the authority over it,” Nadella writes.

His guiding principle comes from cybersecurity: start from a baseline of distrust. AI models should be treated no differently from users on any sensitive network — as potential insider threats subject to strict limitations on what they can access and do. And those restrictions must be enforced by conventional software, not by instructions the models can choose to ignore. The limits on what the machines can do must be strict, auditable and wholly under human control.

Nadella’s approach puts the focus back where it belongs: on companies deploying AI solutions rather than on dreams of regulating intelligence itself. Instead of imagining that government bureaucrats can reliably certify a model as safe, he recognizes that complex systems incorporating AI must themselves be designed to operate securely. AI creates all sorts of new challenges, but engineers have been grappling with the underlying problem for decades. And liability enforces discipline.

Luddites smashing weaving looms failed to stop the Industrial Revolution. Let’s hope El-Sayed’s campaign to halt the development of increasingly intelligent models also doesn’t work. The challenge is to make sure that, however smart the machines become, humans remain firmly in control.