Bias in the Loop: How Human Judgment Shapes AI-assisted Anti-fraud Work

PIA's Executive Director Gavin Ugale discusses downstream cognitive bias in AI-enabled fraud detection for AGA's 2026 Summer Journal.

---

Article at a Glance

AI tools are increasingly used by federal agencies to detect fraud, but deploying them responsibly means accounting for two distinct types of bias: flaws embedded in the AI model itself, and the cognitive tendencies that shape how human examiners interpret and act on AI outputs. These human biases, like over-relying on an AI's risk score or gravitating toward familiar fraud patterns, aren't failures of judgment; they're predictable features of how people make decisions under pressure. The article argues that agencies, oversight bodies, and federal AI policy need to treat human-AI interaction as a design challenge, not an afterthought. That means building pilots, audit frameworks, and governance practices that actively identify and account for how people actually use these tools - not just whether the underlying model is accurate.

Originally published in the AGA 2026 Summer Journal