We just learned that Waymo is recalling about 4,000 self-driving cars because they can’t read highway construction signs.
According to Road & Track, some Waymo cars “failed to recognize and drove past ramp closure signs into pre-planned freeway construction zones; Waymo recorded six of these events in April 2026 in Phoenix, Arizona. In an additional seven incidents in May 2026 in San Francisco, California, the offending Waymos drove between cones designating lane closure.”
That’s a) terrifying b) a ridiculous engineering failure. Construction is hardly a rare condition of highway driving. How did cars not equipped for this get released into the world?
Because someone, somewhere, believes certain sacrifices must be made for progress.
Waymo’s mistake is a very bad example of the kind of real-world exception conditions that have been derailing self-driving cars for more than a decade. Those same pesky, weird realities will soon make many wild promises about artificial intelligence look a equally silly — and cost a lot of investors a lot of money.
A few years ago, I wrote a book called The Plateau Effect with Hugh Thompson, former CEO of Symantec (this piece is my opinion alone, fyi). The concept of the book is simple: plateaus appear everywhere in life, and they bedevil everything and everyone that seems to be making quick progress. People losing weight hit plateaus; people learning guitar or a new language hit plateaus; drugs hit plateaus. Once you start to look, you see plateaus everywhere. A simple example for geeks: Finding the first 99% of bugs in software is relatively easy. Finding the last few is often a nightmare — finding the last 1 percent can take longer than finding the first 99 percent. Sometimes, they’re never found. And so it will be for AI. The Plateau Effect looks like one of those calculus curves where you approach…but never reach…a limit.
Life is full of these kinds of exception conditions. Shit happens. One might say Murphy’s Law suggests that weird stuff like this is normal. Computers don’t do well with irregular inputs, as I’m sure you know. “That….does….not….compute…”
This is part 2 of a three-part miniseries on Artificial Intelligence.
Read part 1: Disarm AI, yes, but the Pope was just getting started
Read part 2: Artificial Intelligence needs a police blotter. Wait, it already has one!
Were a tech mogul sitting here, he’d furiously grab my keyboard at this point and tell me that millions of miles of data will solve this problem — there are only so many exception conditions on the roads, and eventually the supercomputer in the sky will be prepared to deal with all of them. Even without The Plateau Effect, this feels foolishly optimistic to me. Exhibit A is Waymo’s current inability to deal with road construction. But assuming that can be fixed, there will be weird, pesky, unexpected road conditions for as long as humans roam the Earth. Dealing with the first 99% won’t be so hard. Dealing with that last 1 percent will be very, very hard. It’s not going to happen soon, with self-driving cars and many AI applications.
Elon Musk has been promising full self driving cars will be ready in a year or two since…2013. There’s nothing new about this kind of hype cycle. It’s meant to generate excitement and investment so early investors can make off with a lot of cash, funded by later investments. And this pattern keeps working as long as investors believe in tomorrow. Forget autonomous cars; the real future is on Mars! And so it goes.
See, it’s easy to make a car that can ride on highways which can follow predictable patterns — roads you could drive with your eyes closed. But when an 18-wheeler loses a tire tread and cars in front of you start to swerve and dart, well, your eyes better be open. Another example, shared with me recently by Sean McGregor of the AI Incident database: When a dog gets loose in stop-and-go traffic, What Would Waymo Do?
But what about all the dire warnings of AI software teaching itself how to destroy humanity? First off, if you were actively building a tool that would blow up the world, wouldn’t someone at these firms throw the assembly line emergency stop switch? But more to my point here: I suspect AI investors don’t mind all these doomsday scenarios because they help build buzz. I’ve watched entire documentaries about scary AI futures that don’t even raise the question: Are we sure this stuff will actually work? The a priori assumption that we’re successfully building God is pretty valuable to these early investors.
I’m *NOT* saying none of this stuff works. AI is great at writing perfectly trite emails, though my smartphone recently replied “how are you?” to someone on my behalf when I had no intention of asking. That was awkward. Can a Tesla in FSD mode drive better than a teenager playing with his smartphone? Yes, I think so. I drive a car with a lot of automated safety technology that I think works great. But can a Tesla on autopilot deal with a dog on the highway better than a human with good judgment? Not yet. And, I’ll wager, not for a long time.
I’m harping on this point because ignorance of The Plateau Effect sets up a perfect situation for those who benefit from AI hype. Tech moguls and PR firms generate a lot of attention and money with wild promises about tomorrow. At trade shows, I used to call this the ‘next quarter’ game. People would hawk all kinds of great new hardware and software through amazing demos in Q3, then promise it would be ready by Q4. But in tech, Little Orphan Annie is a liar. Tomorrow is much more than a day away.
I don’t care so much about people who want to play this game with investment cash. That’s their business. What I care about is the “sacrifices” that tech firms will decide are acceptable as they run harder and harder into the Plateau Effect. Regulators and voters have to be ready to recognize the dangers we are about to encounter, and the language that will used to distract us. Like Google itself, AI will be in beta mode for a long time.
Be the first to comment