In a recent study at MIT, researchers evaluated facial analysis programs from major technology companies. The research shows major error rates when asked to evaluate women and people of color.
The facial recognition software algorithms were built using neural networks, which look for patterns in a large set of training data. One technology firm boasts 97 percent accuracy in face-recognition software. However, the neural network was trained using a dataset that was 77 % male and 83% white.
MIT researchers evaluated a similar program and used a dataset that included a wider range of people based on gender and skin tones. In the three commercial software systems, error rates of 20.8, 34.5, and 34.7 percent were found for darker-skinned women. This is compared to the 3 percent claimed by one of the system developers.
For more information on these findings, go to: MIT article
I Teach Accounting Information Systems 