Logged-out Icon

Home / Consumer Tech

Age and Racial Biases Found in Autonomous Vehicle Pedestrian Detection AI

Recent research from King's College London reveals disturbing age and racial biases in pedestrian detection AI used in self-driving cars, posing serious safety concerns


The perils of biased artificial intelligence aren’t new, but this week brought fresh concerns into sharp focus. Recent research out of King’s College London reveals significant age and racial biases in the pedestrian detection systems used in autonomous vehicle research. This isn’t just another hurdle for the already complicated world of self-driving cars; it’s a potential road safety nightmare waiting to happen.

Here’s the gist of the study: Researchers put eight widely-used pedestrian detection systems through their paces, using over 8,000 images of people. What they found was unsettling. On average, these systems were nearly 20% more accurate at detecting adults than children. Even more disconcerting, the software demonstrated a 7.5% accuracy advantage for identifying light-skinned pedestrians over those with darker skin.

The root cause? The same culprit we often blame for AI’s discriminatory outcomes: flawed training data. As Dr. Jie Zhang from King’s College explained, the saying “rubbish in, rubbish out” holds true for AI as well. The datasets used to train these systems are awash with lighter-skinned adults, providing a skewed view of who counts as a “pedestrian.” Essentially, the less data there is for the software to learn from about a particular group, the worse it performs when it encounters that group.

But the biases don’t just stop there. The study also found that these inconsistencies are magnified under poor lighting conditions. This means that children and darker-skinned individuals are not just at a disadvantage; they’re at an even greater risk when it’s dark out. It’s especially worrying because car makers, while secretive about their specific software, are generally relying on the same open-source systems scrutinized in this research. That leads us to believe that commercial systems likely suffer from the same biases.

So, what’s the solution? According to Dr. Zhang, we need a two-pronged approach. First off, there needs to be a lot more transparency around how these systems are trained and evaluated. Keeping things under wraps won’t cut it, especially when lives are at stake. Openly sharing performance metrics would allow these systems to be objectively assessed and would put pressure on developers to fix glaring issues.

However, transparency alone isn’t the magic bullet. The AI community, including car manufacturers, needs to ensure that their systems are inclusive and representative of all pedestrians. That means not just diversifying the data these systems are trained on but also advocating for stricter regulations to hold them accountable