On June 25, 2019, as part of their continuing work on the AI Auditing Framework, the UK Information Commissioner’s Office (ICO) published a blog setting out their views on human bias and discrimination in AI systems. The ICO has also called for input on specific questions relating to human bias and discrimination, set out below.

The ICO explains in its blog how flaws in training data can result in algorithms that perpetuate or magnify unfair biases. The ICO identifies three broad approaches to mitigate this risk in machine learning models:

  1. Anti-classification: making sure that algorithms do not make judgments based on protected characteristics such as sex, race or age, or on proxies for protected characteristics (e.g., occupation or post code);
  2. Outcome and error parity: comparing how the model treats different groups. Outcome parity means all groups should have equal numbers of positive and negative outcomes. Error parity means all groups should have equal numbers of errors (such as false positives or negatives). A model is fair if it achieves outcome parity and error parity across members of different protected groups.
  3. Equal calibration: comparing the model’s estimate of the likelihood of an event and the actual frequency of said event for different groups. A model is fair if it is equally calibrated between members of different protected groups.

The guidance stresses the importance of appropriate governance measures to manage the risks of discrimination in AI systems. Organizations may take different approaches depending on the purpose of the algorithm, but they should document the approach adopted from start to finish. The ICO also recommends that organizations adopt clear, effective policies and practices for collecting representative training data to reduce discrimination risk; that organizations’ governing bodies should be involved in approving anti-discrimination approaches; and that organizations continually monitor algorithms by testing them regularly to identify unfair biases. Organizations should also consider using a diverse team when implementing AI systems, which can provide additional perspectives that may help to spot areas of potential discrimination.

The ICO seeks input from industry stakeholders on two questions:

  • If your organisation is already applying measures to detect and prevent discrimination in AI, what measures are you using or have you considered using?
  • In some cases, if an organisation wishes to test the performance of their ML model on different protected groups, it may need access to test data containing labels for protected characteristics. In these cases, what are the best practices for balancing non-discrimination and privacy requirements?

The ICO also continues to seek input from industry on the development of an auditing framework for AI; organizations should contact the ICO if they wish to provide feedback.