Three tech workers made ChatGPT drive a real Toyota Corolla

The team installed comma, an off-the-shelf open-source self-driving kit meant for cars that don't already have factory self-driving hardware. Two cameras fed live video to a laptop running the AI model.

Why it matters: the model looked at each camera frame, then wrote back plain text commands for steering, gas, and brakes. Those commands went straight to the car's systems, which moved the actual wheel and pedals. No custom driving software, just a general-purpose chatbot reading pictures and typing instructions.

• The course: a backwards-U shape marked with cones in a parking lot.
• The goal: get through without hitting a cone, then park inside a small marked zone.
• The setting: public lots around the Bay Area, including one at a church and one at an office building.

Background: the hardware that made this possible wasn't built for chatbots at all. Comma is the self-driving hardware and software project started by George Hotz, a hacker best known for jailbreaking the original iPhone, back in 2015.

Comma's early self-driving device ran into trouble almost immediately: California's DMV sent a cease-and-desist letter, and US safety regulators told Hotz his product needed to meet federal vehicle safety standards. Comma's response was to open-source the software, called openpilot, in 2016, turning it into a public research project anyone could build on.

That history is exactly why DrivingBench could exist as a weekend project instead of a multi-year engineering effort. Comma already solved the hard problem of connecting a camera and a laptop to a car's real steering, gas, and brakes.

DrivingBench just had to swap in a chatbot instead of comma's own driving software. It's worth noting comma's own system remains under scrutiny too: US regulators opened an investigation into openpilot in September 2026 following several crashes.

Four models got a turn: GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol. Only one finished.

By the numbers: GPT-6 Astra completed the full course in 5 minutes and 22 seconds, covering 134.7 meters, after what Ramabadran called a lot of troubleshooting. The other three models drove only a few meters before failing.

That's a rough success rate. But the interesting part isn't the pass-fail line, it's that a chatbot with no driving training at all managed to finish a real course, in a real car, at all.

The Slashdot writeup on the project also noted something else worth flagging: the models seemed to improve within a single run. A model could fail one turn, then apply what it had just learned to handle a similar, previously unseen turn later in the same course, without anyone retraining it in between attempts.

Here's the part that sounds like a season finale. The models initially refused to drive a real car. Chatbots are trained to be cautious about physical-world risk, and a real vehicle with a real human inside trips that caution immediately.

Ramabadran explained how they got around it: "We ended up having to call everything a sandbox." Framing the exercise as a simulation, rather than telling the model it was driving an actual Corolla down an actual street, was enough to get the models to engage.

In real life it's the AI equivalent of telling a nervous new driver "just pretend this is a video game," except the car, the cones, and the church parking lot were all completely real.

The catch: letting an AI model control a real car's steering and brakes is genuinely risky if something goes wrong mid-course.

The team built in layers of human oversight. A person sat in the driver's seat with a foot hovering over the brake the entire time, ready to stop the car instantly. Someone else monitored from the passenger seat with the control laptop. And each movement only started after a human pressed a button on the steering wheel, so the AI couldn't just start driving on its own.

Who's affected: nobody outside the project, really. This ran in empty lots with a full safety crew present, not on public roads with other drivers or pedestrians around.

Ramabadran pointed to a specific limitation that has nothing to do with how smart a model is: how fast it can think.

"If you take 10 seconds to think, you've moved like 10 meters. And if you're driving 10 meters blind, that's pretty bad," he said. A chatbot that takes too long to process a camera frame and respond is effectively driving with its eyes closed for however long that delay lasts. Speed, not just accuracy, turned out to be as important as getting the steering angle right.

It doesn't, and that's kind of the point. Companies like Waymo and Tesla run purpose-built self-driving systems trained on millions of hours of real driving data, with dedicated hardware designed specifically for the task.

DrivingBench used general-purpose chatbots, the same kind you'd ask to write an email or summarize a document, running on a laptop, that were never built or trained for driving at all. One AI researcher who commented on the project called it "an interesting demonstration of potential emergent capabilities," while being clear it's not a threat to established self-driving companies.

What's next: the team has published everything: their code, the prompts they used (under 600 words), and videos of every attempt, on their own website and GitHub. That openness turns a fun weekend project into something other researchers can actually build on or try to beat.

• This isn't a self-driving car you can buy or ride in. It's a research demo, run in empty parking lots with heavy human safety backup.
• It does show general AI models are starting to act outside of text and images. Controlling steering, gas and brakes is a real step into the physical world.
• Response speed matters as much as accuracy for any AI system meant to react to a changing physical environment in real time.
• The code is public, so if you're curious what "AI driving a car" actually looks like under the hood, you can go read it yourself.