Waymo customers woke up Thursday to a jarring change: their robotaxi app suddenly wouldn’t route them on freeways anymore. Trips that normally took minutes were now projected to take much longer on surface streets. The company later confirmed it had suspended freeway driving across all US markets over concerns about construction zones. It also paused operations entirely in Atlanta and San Antonio because its cars can’t reliably handle flooded roads.

This is a big deal, and not for the reasons the skeptics will claim.

The narrative breaks down

For years, the autonomous vehicle pitch has followed a predictable arc. Companies test in controlled environments, then carefully expand to city streets, then confidently announce freeway capability as proof they’re ready for primetime. Freeways are supposed to be easier than city driving: higher speeds but more predictable behavior, fewer edge cases, less chaos.

Waymo already offered freeway service. It worked. Customers used it. And now it’s gone.

The company pulled it back because of construction zones. Not because of some exotic edge case or once-in-a-decade weather event. Construction zones. The thing that exists on virtually every American freeway at any given time.

This isn’t a temporary setback. This is Waymo saying out loud what the entire industry knows but rarely admits: the last 10% of self-driving is vastly harder than the first 90%, and some of that last 10% involves completely mundane situations that humans handle without thinking.

What actually matters here

The reflexive response will be to dunk on Waymo, to say this proves self-driving is a fantasy, that we were promised flying cars and got flashing construction cones instead. That misses the point entirely.

Waymo did the right thing. They identified a scenario where their system wasn’t performing safely and they pulled back. That’s exactly what you want from a company operating autonomous vehicles at commercial scale. The alternative is Tesla’s approach: ship it, claim it works, let the crashes speak for themselves.

The construction zone problem is genuinely hard. Lanes shift. Signage is temporary and inconsistent. Human workers are present. The rules that normally govern a freeway suddenly don’t apply, and you need to navigate based on ambiguous visual cues and assumptions about what other drivers will do. This is precisely the kind of scenario where AI struggles: high stakes, incomplete information, rapid contextual changes.

The flooded roads issue in Atlanta and San Antonio is even more revealing. This isn’t a bug in the software. It’s a fundamental limitation of the sensor suite and decision-making system. How deep is the water? Is it safe to proceed? Should we route around it? Humans make these judgments all the time, often badly, but they make them. Waymo’s system apparently can’t make them reliably enough to keep operating.

The honest version of self-driving

What Waymo is demonstrating, perhaps unintentionally, is what a mature autonomous vehicle deployment actually looks like. It’s not “full self-driving everywhere.” It’s a carefully scoped service that operates where it works and pulls back where it doesn’t.

That’s fine. That’s probably the right model for the next decade at least. Robotaxis on surface streets in good weather in well-mapped cities. Expand gradually. Pull back when you hit limitations. Don’t overpromise.

But it also means we need to stop pretending we’re on the verge of a transportation revolution that will eliminate car ownership and reshape cities. We’re not. We’re in the early stages of a long, grinding process of figuring out which specific use cases autonomous vehicles can actually handle safely.

Waymo has 150,000 paid rides per week across its markets. That’s real commercial traction. It’s also a tiny fraction of total ride-hailing demand, let alone total driving. And now we know that freeway driving, which everyone assumed was a solved problem, is back off the table.

The question isn’t whether Waymo will fix the construction zone issue. They probably will, eventually. The question is what else is lurking in that last 10%. What other mundane scenarios will turn out to be surprisingly hard? How many times will we see features launched and then quietly pulled back?

The self-driving industry has spent 15 years telling us complexity is just a software problem. Waymo just admitted, in the most concrete way possible, that some complexity doesn’t yield to more training data and better models. Sometimes you just have to tell customers their freeway trip isn’t available today.

That’s not failure. That’s reality. We should see more of it.