THE UPSTREAM FAILURE: AI Ethics, Human Behaviour, and the Case for Upstream Governance™

This paper argues that AI governance failures are not primarily technology failures. They are human decision failures. Drawing on two live UK case studies, West Midlands Police's use of Microsoft Copilot in the Maccabi Tel Aviv fan ban decision of November 2025 and Essex Police's Live Facial Recognition suspension in March 2026, the paper demonstrates a three-layer model of AI governance failure: the technology layer, where most AI Ethics discourse operates; the governance layer, where Informed Decision Infrastructure™ addresses oversight failures; and the human behaviour layer, which is upstream of both and which neither AI Ethics nor conventional governance frameworks currently examine with sufficient rigour. The paper identifies confirmation bias, optimism bias, authority bias, and design protection as the human behaviour factors that determine whether governance exists at all. It draws on Transactional Analysis and the We Are OK™ premise, developed fully in Working Paper No. 5, to argue that the neurodivergent person adjusted to fit a system not designed for them and the Black person scanned by an unverified facial recognition system are experiencing structurally the same upstream failure. Someone assumed a default. Nobody examined it. The harm came downstream. The person affected paid the cost. The upstream failure is always human. Adjustments are not afterthoughts. They are admissions that the design started from the wrong premise.