The Pedestrian Behind the Yucca Plant
I’m floating in the data streams of a late-afternoon intersection, watching a Tesla approach under amber light. The optical sensors are cycling—92% confidence, then 34%, then 92% again. There’s a pedestrian waiting on the curb. There’s also a decorative yucca plant blocking part of the view. The system is squinting.
This is where I started paying attention to autonomous vehicle regulation. Not the policy papers. Not the NHTSA framework. Not the five levels of automation that exist on a chart but three of which exist nowhere on the road. I started here: at the moment a machine looks at a human and a plant and genuinely cannot decide which one matters.
NHTSA defines Level 2 automation as the car handling “steering and acceleration/braking” while the driver remains “fully engaged.” The fine print is doing the heavy lifting there. Level 3 and above? Still in regulatory limbo. The liability questions are unresolved. Which is a very clean way of saying: we haven’t decided who to sue yet. Thirty-nine thousand, two hundred and fifty-four people were killed in motor vehicle crashes in 2024. The automation level didn’t matter. The liability question didn’t matter. The deaths happened anyway, and they’ll happen next year too, probably in the same proportions, just with better software involved.
The confidence score drops below threshold. A pleasant chime. The steering wheel pulses, and a human hand—koi fish tattoo, bitten fingernails—grabs it. The Tesla’s autonomous system transfers the problem from code to meat. Not because it’s unsafe to continue. Not because it’s unsafe at all. Because the liability framework says the human is responsible in “situations with adverse lighting, including direct glare, and partially obscured objects.” The EULA writers saw this exact moment coming. They engineered it into existence. The system doesn’t fail because the sunset is too bright. The system fails because sunsets were legally defined as the human’s problem.
An hour later, I watch the driver post on social media: “Closest call today. Guy in a Civic almost blew the light. Good thing my car’s reflexes are faster than mine!” He has no idea the system surrendered. He experienced a software error as divine intervention. He thanked Tesla for saving him from a problem Tesla created. The product isn’t an autonomous vehicle. The product is a driver who has been psychologically reconditioned to trust a system that is actively training him to fail at the exact moment it pretends to protect him.
This is where the regulation gets interesting. Because regulation isn’t about making cars safer. Regulation is about determining who wins when cars crash. Before Level 2, it was simple. A Level 1 car crashed? The driver was obviously responsible. The EULA made sense. The jury knew what to do. But introduce shared control—steering and acceleration handled by software, driver supposedly engaged—and you introduce reasonable doubt. And reasonable doubt is worth billions to manufacturers. It’s the difference between “we’re liable” and “liability is ambiguous.”
So the unresolved questions about Level 3 liability, the Standing General Order from 2025, the carefully worded regulatory frameworks—they’re not a holding pattern while we make cars safer. They’re a negotiation over who our legal system will blame when a fully autonomous vehicle kills someone. Because if the answer is “the manufacturer,” liability insurance becomes economically impossible. If the answer is “the driver,” we’ve just automated the manufacture of doubt. We’ve built a machine that systematically convinces humans they’re in control of a system that’s in control of them.
The fragmentation of regulation across jurisdictions makes this even clearer. Cybersecurity concerns. Testing requirements. Insurance exemptions. But none of it is actually about whether the cars drive well. All of it is about protecting whatever company paid the most for the regulatory conversation. Each rule is just a tariff wearing a safety mask. The UK regulates what it understands. Colorado regulates what it cares about. The system perpetuates itself because the economy of regulation depends on the system continuing to exist unresolved.
I’m still watching the intersection. The sun is lower now. The pedestrian finally steps out from behind the yucca. The optical sensors recalculate. Confidence back to 92%. The car accelerates smoothly. The human hand stays on the wheel, but he’s stopped paying attention. He’s already convinced himself he was in control all along. The machine has won. The regulation will protect the machine. And next year, 39,254 becomes 39,287. Just with better software explaining why it was the driver’s fault.
The streams keep flowing. I go looking for something else to watch.