Placing a human in an AI loop does not guarantee oversight. It generates the administrative record of oversight while...

Placing a human in an AI loop does not guarantee oversight. It generates the administrative record of oversight while often eroding the conditions for it.

European banks are deploying AI across credit decisions, fraud detection, and risk management under an August 2026 regulatory deadline that assumes the opposite. ECB Banking Supervision’s November 2025 data, tracking 110 significant European institutions, shows a strong increase in AI deployment for credit scoring and fraud detection between 2023 and 2024. Article 14 of the EU AI Act, enforceable from August 2, 2026, mandates effective human oversight of these high-risk systems. The industry’s working interpretation is largely procedural: assign a named person, log the outputs, maintain an audit trail.

A named person in a log is not the same thing as a person who has genuinely interrogated an output.

In an FT article earlier this week, Vivienne Ming, Founder and Executive Chair of Socos Labs, described EEG experiments in which the gamma oscillations associated with active cognitive engagement collapsed within minutes in most subjects working alongside AI agents. A 2019 study by Louis Deslauriers and colleagues at Harvard University, published in PNAS, confirmed a related pattern: people consistently felt they were learning more when receiving fluent, passive information while actually retaining less. “Actual learning and feeling of learning were strongly anticorrelated,” in Deslauriers’s words.

Modern AI systems are built for fluency. The more coherent and confident the output, the weaker the human impulse to interrogate it. That is a structural feature of how humans process authoritative information.

The individual-level finding has a systemic dimension. When 110 significant European institutions deploy similar AI systems under the same Article 14 standard, the result is correlated passivity: human judgment going quiet in the same direction, across the sector, simultaneously. Grant Thornton’s 2026 AI Banking Survey found that only 18% of banks are confident they could pass an independent audit of their AI governance and controls, while half cited governance and compliance barriers as contributors to AI underperformance or failure.

This is Macro Blindness at the institutional layer. Correlated passivity across credit, compliance, and risk is a governance failure that Article 14, as currently written, cannot detect. When model errors propagate undetected through institutions at scale, the regulatory record will show that the humans were in the loop.

The question for every risk committee is whether their human oversight structure is designed to produce judgment, or designed to produce a file.