Towards causal benchmarking of bias in face analysis algorithms

Synthetic images reveal true algorithmic bias: Current methods for measuring bias in facial recognition systems use real datasets, but this approach conflates algorithmic bias (flaws in the AI system itself) with dataset bias (imbalances in training data). The researchers developed a method using synthetic images that systematically vary specific attributes like gender and skin tone to isolate what the algorithm itself is doing wrong.

Human perception is essential for fair AI testing: The method relies on human observers both to guide how images are manipulated and to evaluate algorithmic bias, ensuring that the synthetic test images are realistic and meaningful rather than based solely on technical assumptions about what matters.

Synthetic testing is more equitable and practical: Unlike observational datasets, synthetic transects can evenly sample minority and intersectional groups (e.g., women with darker skin tones), making it possible to detect bias affecting smaller populations that might be overlooked in traditional testing. This approach also reduces ethical concerns about using real people's data and is faster and cheaper to implement.

Better bias testing can become widely accessible: By making bias measurement more affordable and efficient, this method could enable smaller organizations and researchers to regularly test their face analysis systems, potentially democratizing responsible AI practices beyond well-resourced tech companies.