AI was trusted with a real Toyota for the first time – what came of it

An American engineer connected ChatGPT, Claude, and Grok to the control systems of a Toyota Corolla and conducted tests on an empty parking lot. The main problem was a delay in decision-making: the neural network would determine the need for a maneuver, but by the time the command was executed, the car had already deviated from the trajectory. The models also struggled to estimate distances to objects and issued incompatible commands, such as simultaneously demanding full power and full braking.

An American engineer conducted an experiment by connecting three leading neural networks—ChatGPT, Claude, and Grok—to the steering, throttle, and brake systems of a real Toyota Corolla. The tests took place on an empty parking lot: cameras transmitted images of the surroundings to the models, after which the neural networks issued text commands specifying the steering angle, throttle position, and braking intensity. The main problem was a delay in decision-making. The neural network could determine that the car needed to change direction, but by the time the command was executed, the vehicle had already significantly deviated from the intended trajectory. The models also had difficulty estimating distances to objects due to a lack of full 3D perception. In some cases, the neural networks issued incompatible commands—for example, simultaneously using full engine power and fully braking. Different models showed different results. GPT-6 Astra handled control better but lacked reaction speed. Grok acted faster but had issues with accuracy and safety. Claude recognized potential dangers well but reacted too slowly and cautiously. In a separate test, GPT-6 Astra completed the route in 5 minutes 22 seconds, Claude Fable 5.1 covered 45% of the distance, Grok 4.6 stopped at 11%, and GPT-5.6 Sol covered only 6% of the route.