Discriminatory Algorithms

In this presentation, the Centre for AI & Digital Ethics (CAIDE) gives an introduction to how machine learning techniques learn to make decisions that discriminate against minorities, and discuss why removing this discriminatory bias is difficult. Machine learning technologies are powerful techniques for discovering patterns in data, which can be used to make new inferences. However, recent research shows that these techniques can be unreasonably biased against minority groups; for example, hiring algorithms that exclude women, and recidivism algorithms that assign higher risk scores to people who aren’t white. Such algorithms can operate at a scale not possible for human decisions makers, meaning the impactful decisions that are harmful to people can be widespread. In this interactive presentation, we will lead participants through a series of short exercises that demonstrate the basics of machine learning, how and why machine learning techniques learn discriminatory biases, how some biases can be discovered and mitigated, and why mitigation techniques do not entirely solve the problem.