Sometimes it seems like we’ve been talking about self-driving vehicles our whole lives.
The idea that you can put an LLM in a car, equip the vehicle with sensors and cameras, and set rules for safe operation has been around for a while. People who live in Palo Alto are probably used to seeing a driverless car quietly humming around on the street, but in middle America, many have never seen this kind of thing in practice. In other words, self-driving car are still rare.
So why has it taken so long?
The answer seems to come down to liability.
I read this recent article that attempts to explain why self-driving car implementation has hit a snag. Keep in mind some of these cars are already on the streets: the reporting shows that Waymo autonomous vehicles are involved in crashes a full 68% less than human-driven conveyances.
The words “trial lawyers” appear in the headline. That’s your first clue.
It turns out that parties involved in the legal system can’t agree on what kind of regulation to impose on self-driving cars. Should it be state? Federal? Who do you blame if something goes wrong?
“Supporters say autonomous driving technology could eventually prevent thousands of crashes,” writes Jessica Wong for Moneywise. “But efforts to speed its rollout through federal legislation have run into resistance from trial lawyers, who argue the proposed rules don’t go far enough to protect consumers or hold manufacturers accountable when autonomous vehicles do happen to fail.”
Reading this, you start to get an idea of why we don’t see autonomous cars in dealerships across the country, replacing those antiquated old buggies that you have to drive yourself.
Sophistication in Design
I wanted to go over some things I heard in a talk at a recent TedX Boston event this year. Jason Hansberger, director of the DAF-Stanford AI Studio, talked about his experience in the defense industry, and put forward a novel suggestion involving two different modes of thinking about autonomous design.
In the Air and On Land
One of Hansberger’s arguments is that in designing autonomous systems, one should start with a more controlled scenario rather than a wilder one. He described an op-ed he wrote after having this epiphany while working in aviation:
“If we are going to automate transportation, the wide open, chaotic road won’t be the first place,” he said. “The mismatch is too great. The favorable operating environment and pre-existing autonomy in aviation meant it was going to be the highly procedural cockpit. Pilots, I argued, before drivers.”
But the gist of his thoughts on self-driving cars on roads has to do with two design theories he calls Layout Replace and Layout Maximize. I’ll use Hansberger’s own words to describe these:
“Layout Replace assumes the algorithm should do the entire job, start to finish,” Hansberger said. “No assistance, no interruption. If the system needs a human, the system has failed. Replace is the assumption behind viral demos and marketing: the robot that folds the laundry by itself, the drone that delivers your package by itself, the car that drives itself. Layout Maximize assumes something different. It assumes the algorithm doesn’t have to do everything alone. It assumes the algorithm’s most important capability isn’t doing the task alone. It’s knowing when to ask for help.”
Here’s a more concise version of the above, reiterated by Hansberger:
“Replace says for every part of the task, the algorithm can do it. Maximize says there exists at least one part of the task that the algorithm cannot do alone.”
Evaluating Choices
Hansberger pointed to some flaws in the Replace principle:
“The moment I pick replace as my design assumption, I’ve committed to a specific claim about the world,” he said. “A claim that is almost never true, because in the real world, there are many categories of constraints that don’t disappear when the model gets bigger.”
That gap, he noted, is only part of why it’s problematic to go with a Replace paradigm. The other big issue is that humans don’t like to be replaced.
“Users don’t trust a black box in full control of certain things, and even when they do, regulation may mandate a human in the loop,” Hansberger said. “And practically speaking, the marginal cost of increased automation often doesn’t outweigh the benefit. Most visibly, people mobilize when they think AI is coming for their jobs. … just this month, there were stories of graduates booing commencement speakers who brought up AI. Different industries. Different generations. Same floor, as any one of these is enough to make pure Replace impossible.”
Explaining how “every constraint becomes a floor,” Hansberger brought the audience through various charts showing projected outcomes, also supporting the idea that there should be a human in the loop.
More Reasons
Promoting the hybrid principle of Maximize, Hansberger talked about some of the failures of Cruise robo-taxis, and had this to say about a set of trails where AI assisted pilots:
“The AI added all of its value in the moments where human cognition hit its limits: novel situations, incomplete information, and judgment calls. That is not a replacement. That’s a dual. That’s what Layout Maximize looks like in the cockpit. Here’s where autonomy stops being binary and starts being a loop. The AI asks for help at its boundary. The constraint is named. The team—engineers, policymakers, and operators—solves for that specific limit, and the boundary moves. This is how autonomy actually grows in the real world.”
Matriculation in the Age of AI
Here’s what Hansberger had to say to new grads:
“Anyone designing machines to come and replace you will hit a floor of performance before they hit a threshold of success. Unfortunately for our recent graduates, that could be after you’ve been fired, or never hired in the first place. However, AI that sits next to you as your dual—that’s different. Find the people building that. Go work for them, because that’s the frontier.”
As for designers – he concluded by saying this:
“The useful question was never: when will AI replace us?” he said. “The critical question is whether we are going to design AI to replace human cognition, or to maximize it. I think we should choose Maximize.”
That’s a pretty clear argument, and I thought that it holds water in 2026. Why should we keep knocking our heads against the wall, looking for those diminishing returns, when the whole thing can be solved by just a little bit of human oversight or potential intervention? We’ve already solved the technology problem. We need to solve the responsibility problems, the social problems, and the problems around risk and liability. Nothing is perfect, as the conductor of potentially lethal vehicles that, in some ways, defy physics. Humans certainly aren’t. So if you’re waiting for a car that just “drives itself” while you sleep or have a sandwich in the back seat, you might be waiting a little longer.
