Engineers tested GPT, Claude, and Grok to drive a real car

In short: Three AI engineers tried letting GPT, Claude, and Grok control a real car, and only one system completed the task.

What happened

A WIRED report describes an experiment by three AI engineers at the startup Axiom. They wanted to see if today’s popular chatbots could handle a physical job, not just answer questions on a screen.

They put three systems, GPT, Claude, and Grok, in charge of driving a real car. In simple terms, they tried to turn each chatbot into the “driver” by having it decide what to do next, like steering, braking, and moving through traffic.

Only one of the three was successful. The other attempts did not manage the full driving task. The story highlights how big the gap can be between talking about driving and actually doing it safely in the real world.

Why it matters

It is easy to assume that if an AI can write a good answer, it can also run real things like cars. Driving is different because mistakes have immediate consequences. It is less like writing an email and more like juggling while riding a bike (you have to keep adjusting in real time, and small errors can turn serious fast).

Experiments like this help show what current AI can and cannot do outside a chat window. They also raise questions about testing and safety before anyone treats general chatbots as reliable controllers for machines people depend on.

Source: Wired