Should governments regulate AI like drugs or like cars?

Here's a question I get from policymakers.

Should we approve AI models before they're released, the way we approve medicines? Or set rules of the road and let people build?

The history of both approaches has an answer.

What's the pattern?

Drugs got pre-approval in 1938, after a medicine called Elixir Sulfanilamide killed more than a hundred people, many of them children. The rule: prove it's safe before you sell it. It works because a drug is one specific thing with one specific use, and you can test it.

Cars took a different road. For sixty years there were almost no federal rules. Then in 1965, Ralph Nader published Unsafe at Any Speed, and in 1966 Congress created the agency that gave us crash standards, seatbelts, and eventually airbags. Nobody has to approve your car before you build it. It has to meet standards, and if it fails, someone is liable.

AI is a car, not a drug. It's general purpose. It's built by thousands of actors. It's updated constantly. A pre-approval system would either grind to a halt or approve things that changed the day after. The car model, standards, testing, and liability, scales.

And notice what actually made cars safe. Not the ban some people wanted. The seatbelt, the crash test, and the lawyer.

What should leaders do now?

So, three things, if you sit anywhere near policy.

Set standards for the use, not the model. A model that writes poems and a model that decides parole are the same technology and wildly different risks. Regulate the parole.

Mandate testing you can see. Crash tests worked because the results were public. Require the equivalent for high-stakes AI, and publish it.

And make liability clear before the crash. The single most powerful rule in car safety was that the manufacturer could be sued. Say who's liable when an AI system harms someone. Everything else follows.

We didn't ban cars. We made them survivable. That's the job here.

If an AI system in your jurisdiction harmed someone tomorrow, who would be liable, and would they know it?

Related: Who is accountable when your bank's AI says no?

Sam Rad, The Change Futurist. Keynote speaker on change, transformation, resilience, and AI adoption. Author of Radical Next. Book a keynote