Racial Bias in Detection: The AI systems showed a significant bias in detecting pedestrians based on skin tone. Darker-skinned individuals were identified with 7.52% less accuracy compared to their lighter-skinned counterparts. This disparity was even more pronounced in low-contrast or low-brightness settings, such as at night.
Age-Related Bias: The study also found that these AI systems were 20% less likely to detect children compared to adults, indicating an age-related bias.
Methodology of the Study:
Researchers annotated a total of 8,111 images with various labels, including gender, age, and skin tone, to assess the AI systems’ accuracy in pedestrian detection.
Implications for Autonomous Vehicles:
While the specific AI models tested in the study were not those used by driverless car companies, the researchers believe that the industry’s models are likely built upon similar open-source models and therefore may exhibit similar biases.
Jie Zhang, a co-author of the study, points out the secrecy around the AI models used in autonomous vehicles, classifying them as proprietary information. However, the similarities with open-source models suggest that these commercial systems might also have inherent biases.
Broader Concerns About Machine Bias:
The study adds to the growing body of evidence that machine learning algorithms can perpetuate and amplify existing societal biases, particularly those based on race and age.
With the increasing integration of AI in everyday life and critical systems like autonomous vehicles, these biases could lead to severe consequences, including endangering lives.
The Need for Immediate Action:
The findings underscore the urgent need for regulatory measures and a proactive approach to address these biases in AI systems, especially those used in safety-critical applications like driverless cars.
Waiting for regulatory catch-up post-tragedies is not a viable option, emphasizing the importance of preemptive action to ensure AI systems are fair and safe for all.
In conclusion, the study from King’s College London serves as a crucial reminder of the challenges in developing unbiased AI systems. As autonomous technology continues to advance and become more prevalent, ensuring these systems are equitable and safe for all users, regardless of race or age, is of paramount importance.
