AI safety sensors in self-driving cars are less accurate for people of color and children

The explosion of artificial intelligence (AI) has brought a clear trend: Bias in training machine learning systems leads to discriminatory behavior in the real world .

A recent study by a group of scientists in the UK and China shows that pedestrian detection software equipped on self-driving vehicles - which is being used by many car manufacturers - may be less effective when the subject is people of color or children, leading to these people being at greater risk when participating in traffic.

AI in autonomous vehicles tends to be less effective with people of color and children, potentially leading to serious accidents.

Although gender showed only small differences in accuracy, the researchers found that the AI system was less accurate in recognizing pedestrians with darker skin tones.

“Minority individuals who have been denied many important services are now facing serious accident risks,” said Jie Zhang, a computer scientist at King’s College London and a member of the research team.

Specifically, the report noted that the system was 19.67% more likely to detect adults than children, and 7.52% less likely to detect those with darker skin tones.

“Overall, this research sheds light on the fairness issues facing current pedestrian detection systems, highlighting the importance of addressing age- and skin-related biases,” the team wrote in their paper. “The insights gained could pave the way for more fair and unbiased autonomous driving systems in the future.”

According to members of the research team, their report did not use the exact software that companies like Tesla install in their cars, as this is a trade secret, instead the research was based on similar open-source AIs used by businesses.

“It is essential that policymakers enact laws and regulations that protect the rights of all individuals and appropriately address these concerns,” the team urged lawmakers to regulate self-driving car software to prevent bias in their detection systems.

(According to Insider)

Source