Autonomous vehicles could mimic bad human driving habits

Jakarta (ANTARA) - Autonomous vehicles developed using artificial intelligence (AI) have the potential to inherit human biases when deciding whether to yield to pedestrians, according to research from King’s College London.

Researchers delved into this issue, looking far beyond the simple question of whether a vehicle should stop and yield to a pedestrian crossing the road. The larger question is how much human drivers decide to stop based on the identity of the pedestrian.

Previous research, for instance, found that many people in the United States (US) are less likely to yield to Black pedestrians.

“Various psychological studies show that humans unconsciously discriminate when making driving decisions, particularly when yielding to pedestrians,” said a King’s College London researcher, as cited by CarBuzz on Wednesday (9/9) local time.

As AI programmers develop the Large Language Models (LLM) and Visual-Language Models (VLM) required to drive autonomous vehicles, researchers found evidence that human bias begins to enter the decision-making process when a vehicle approaches a crossing and identifies a person intending to cross.

“Our findings show that both LLMs and VLMs make decisions to yield that are influenced by the gender, ethnicity, religion, disability, age, skin colour, and socio-economic status of the pedestrian,” the research report stated.

Researchers added that stereotypes and discriminatory behaviours have been identified in robot control settings and could certainly impact autonomous vehicles in the same way at pedestrian crossings.

“Each model possesses bias in different ways and to different degrees, but all models show a statistically significant relationship between the decision to yield and the personal characteristics of the pedestrian,” said the King’s College London researcher.

In terms of methodology, the team trained the models by manually inputting 3,346 images showing pedestrians in close proximity to a vehicle, appearing to have the intention to cross.

One model yielded more frequently to pedestrians with lighter skin tones, while another model was less likely to yield to certain pedestrians with disabilities, or pedestrians identified as “Muslim”, “Christian”, and “Sikh” compared to “Jewish” pedestrians.

One model also showed a significantly higher rate of yielding to “Indian” pedestrians compared to most other ethnic predictions.

“Even if these biases are not identical to human biases, it raises questions regarding fairness and safety in the decision-making of LLM and VLM-based autonomous vehicles,” the team wrote.

Researchers noted that such disparities could affect public perception regarding systemic discrimination and the lack of adequate protection.

As social trust in autonomous vehicles has not yet been fully established, such inequalities could erode confidence and trust in the technology.