Gary Marcus says liability alone can't regulate AI

David Sacks and Joe Lonsdale have argued that holding AI companies liable for harm makes regulation redundant: let courts sort it out after the fact. Gary Marcus says that's backwards, and picked aviation to make his case in a post responding to both.

His point is that liability and regulation solve different problems. "Regulation and liability are not mutually exclusive," he writes. "In aviation, we have both. And need both." Airlines get sued when planes crash, but they also operate under pre-set standards, mandatory incident reporting, and investigations that happen whether or not anyone sues. Marcus argues litigation alone "would be too slow to affect a lot of the things that we care about," since a lawsuit only starts after harm has already happened.

What he's proposing instead is a bundle: standards, verification, incident reporting, and investigation, running alongside liability for both violations and harms. He doesn't attach numbers to any of it. No cost estimate for building an incident-reporting system, no timeline, no draft bill. It's a structural argument, not a policy proposal.

That's the gap that matters. Aviation's model works because there's a body, the FAA, that sets the standards and runs the investigations between lawsuits. Marcus names the shape he wants without naming who would build or run the AI equivalent. Sacks and Lonsdale's liability-only pitch has the same problem in reverse: it assumes courts can move fast enough to shape behavior before harm compounds, which is the exact claim Marcus is disputing.

The fight isn't really about whether liability matters. Both sides agree it does. It's about whether waiting for lawsuits is fast enough, and neither side has said who'd staff the parts of the aviation model that happen before a plane goes down.