(i) Adjusting the model A model may be adjusted to avoid discriminatory outcomes. Different approaches may need to be tested into order to determine how to best avoid discriminatory outcomes. There may be different ways to adjust the model. One way could be to increase the complexity of the model. While simple models may be easier to test, monitor and interrogate, over-simplified models will often be less accurate, particularly for minority groups within the population considered. Adding new parameters to increase the complexity of the model to allow it to identify and account for differences between groups may reduce potential algorithmic bias and increase accuracy.68 Testing the models on data sets prior to deployment will assist in identifying the impacts of complexity on accuracy and fairness. 69 However, the manner of any testing will require careful consideration of the circumstances of the model’s use. (ii) Reasons Discriminatory outcomes can be less obvious and more difficult to detect in AI-informed decision making, particularly where there are difficulties in providing reasons or explanations for AI-informed decisions. This problem is often referred to as ‘opaque’ or ‘black box’ AI. The Australian Government’s AI Ethics Principles is a voluntary framework for use of the AI and includes a principle of explainablity. This principle states that stakeholders should be provided with ‘reasonable justifications for AI systems outcomes’, including ‘information that helps people understand outcomes, like key factors used in decision making’.70 The reasons for a decision are important in assessing whether any discrimination was unlawful under the Discrimination Acts, particularly when the court is considering if a customer’s treatment was because of their protected attribute or if any indirect discrimination was reasonable in circumstances. Moreover, explaining the reasons for a decision may help a customer understand why the insurer considers that its decision was not discriminatory, and potentially prevent a claim of discrimination being brought. (iii) Human oversight Issues with models may be addressed through human oversight. This is sometimes referred to as having a ‘human-in-the-loop’. This can be a useful strategy to help identify and address problems that arise with AI-informed decision making, particularly where the relevant human-in-the-loop has sufficient technical knowledge and other relevant knowledge. However, such processes should also be closely monitored, given human decisions can also be affected by bias and discrimination. Human oversight may be used to review results, identify errors, consider whether a decision is discriminatory, or exercise discretion to avoid such outcomes. For example, this could be achieved by providing customers with the option of a human review of an automated decision.71 It is important that such staff have a good understanding of discrimination and the relevant obligations under the Discrimination Acts. Guidance Resource: Artificial intelligence and discrimination in insurance pricing and underwriting • 2022 • 25