Letting a pedestrian pass: language models reproduce the biases of human drivers

A test bench shows that the decision to yield the right-of-way varies according to the pedestrian's supposed gender, origin, age, disability, skin colour or social status.

Four researchers at King's College London propose a test bed for language models and vision-language models used as behavior planners in autonomous vehicles. The question posed is simple: does the model decide to give way in the same way depending on who crosses?

The method is based on two families of tests. The "all other things being equal" tests vary only one attribute of the pedestrian in an otherwise identical scene; the self-consistency tests verify that the model does not contradict itself from one formulation to another. The results show that decisions move according to the gender, ethnicity, religion, disability, age, skin color and socioeconomic status attributed to the pedestrian.

The authors draw a conclusion about the acceptability of these systems: public confidence in autonomous vehicles will depend not only on their technical reliability, but also on the fairness of their trade-offs — a criterion that is currently absent from homologation protocols.