From the course: Introduction to Auditing AI Systems

Types of model audits

From the course: Introduction to Auditing AI Systems

Start my 1-month free trial Buy for my team

Types of model audits

“

- [Instructor] Model audits should always be considered in context of their target principles. Teams should decide if they're auditing for bias, fairness, or compliance. We have to decide if we want to measure system outcomes or specific ML models. We also have to consider what metrics matter, which demographic groups and intersections of them are of interest, as well as whether we're auditing a system before or after deployment. With all this to consider, every auditor must determine what's in scope for their audit. In practice, auditors find themselves limiting scope for various reasons. That can include lack of data or system access, limited compute to make new predictions, or because some regulations don't require comprehensive audits. Data audits analyze how data is collected, manipulated, and stored, in addition to its underlying features, such as protected classes and their proxies. Often, these audits benefit from the use of data governance tools. Data audits can help identify bias in training data and how the assumptions made about that data can skew model predictions. A model audit investigates a single ML model for bias, compliance, or alignment with company policy. This typically involves calculating disparity metrics for model predictions and providing necessary retraining or model updates to fix any identified disparities. A comprehensive system audit involves examining the algorithmic choice, inputs, outputs, and contexts from a sociotechnical lens. It aims to identify issues that a model audit or data audit might miss. In a recent example, a system audit might have identified the use of language models for translation in high-risk applications, such as those seeking asylum, as inappropriate. A system audit looks at the technical details, as well as social contexts to decide if systems are appropriate as deployed. Continuous audits try to test a system's outcomes over time. Part of continuous auditing requires inspecting predictions on new data, but audit results should also be compared to prior audits. This allows teams to make crucial decisions, like to keep models in production or depreciate them.

Contents