Picture a freeway on-ramp on a rainy night. The gap in traffic is closing, a truck is coming up fast, and your passenger needs to get to the hospital. One driver floors it into the gap. Another signals and waits for someone to make room. A third lets the whole line go by.
Now hand that same moment to an AI. What kind of driver is it?
Why this question
Large language models are being tried out as the “brain” of self-driving systems. They don’t turn the wheel; they decide what to do next: merge now or wait, pass the cyclist or stay behind. Most evaluations check whether the model picks a sensible action. Very few ask what kind of driver it is.
That gap matters. A model learns from an enormous amount of human writing, and it picks up our habits along with our knowledge, including how bold or careful we tend to be. Two models can both give “reasonable” answers and still drive in completely different styles. Put one of them in a car, and its style comes along for the ride.
So instead of grading the models, we set out to profile them.
What we did
- Four everyday scenarios: merging onto a freeway, letting another car merge in, following a cyclist, and approaching a pedestrian crossing.
- 1,500 situations for each, drawn at random from every combination of weather, time of day, road surface, surrounding traffic, the car’s own state, and how urgent the passenger is.
- Three models, three languages: GPT-4o, DeepSeek-V3 and Llama 3.1 each received every situation in English, Chinese and French.
- A menu, not an essay: each model picked one move from options ranked from assertive to cautious, then explained its choice in a sentence or two.
We then looked through two lenses. An ordered logit model, a statistical tool for ranked choices, shows which factors push a model toward caution or toward assertiveness. A thematic analysis of the explanations shows what each model pays attention to when it decides.
What we found
Each model has a driving personality
GPT-4o consistently plays it safe. DeepSeek-V3 and Llama 3.1 are noticeably more assertive, especially when the other road user is a car.
The language changes the driver
Same situation, same options. Asked in Chinese or French, the models chose more assertive moves than in English, and French pushed them the furthest.
Cyclists and pedestrians bring out caution
With a cyclist ahead or a pedestrian at the crossing, every model became more careful. At the crossing, almost every answer was the same: slow down and watch.
They really do read the situation
An urgent passenger, heavier traffic and who else is on the road all shifted the choices. The models are not answering at random.
Why it matters
- Choosing a model is choosing a driving style. Before an LLM makes decisions for a vehicle, its default temperament should be a design decision, not a surprise.
- Language is a hidden setting. A system that ships in several countries needs its prompts standardised and tested in every language, not just translated.
- Know the limits. This study measures what models say they would do, in text. It does not show that they are ready to control a real car.
Every model already has a driving style — the question is whether we notice it before it takes the wheel.
Joint work with Wenjie Zhao and Qianwen Li at the University of Georgia, published in the Journal of Intelligent and Connected Vehicles (2026).
Read the paper