AI needs human checks
Regulatory and supervisory guidance is emphasizing formal AI model validation and human‑in‑the‑loop checkpoints for finance — document assumptions, limitations, and validation steps for any AI‑generated output. (mondaq.com)
A Cahill client alert dated March 23, 2026 highlights that generative AI’s nondeterministic outputs challenge traditional repeatable benchmark testing used in model validation. (cahill.com) Mondaq republished that memorandum this week, noting broker‑dealers and investment advisers face strained validation approaches when AI is embedded in compliance, client communications, trading, or research workflows. (mondaq.com) The Federal Reserve’s SR 11‑7 guidance — the baseline U.S. model‑risk framework issued in 2011 — continues to require independent validation, documented model inventories, and a three‑lines‑of‑defense governance structure for models. (federalreserve.gov) The FDIC’s “Supervisory Guidance on Model Risk Management” expands the scope of expectations and states the guidance generally does not apply to FDIC‑supervised institutions under $1 billion in assets unless their model use is material. (fdic.gov) FINRA’s thematic guidance lists model risk management, data governance, customer privacy, supervisory controls, and explicit checks of inputs, parameters, and output explainability as areas member firms must update for AI deployments. (finra.org) The U.S. Government Accountability Office reported in May 2025 that federal regulators are conducting AI‑focused examinations while continuing to rely on existing laws and supervisory guidance and are considering refinements or additional rules. (gao.gov) Market advisers and Big Four firms recommend practical validation steps now expected by supervisors: adversarial/robustness testing, continuous data‑drift monitoring, disparate‑impact/bias testing, retained audit trails, vendor oversight, and defined human‑approval thresholds for material automated actions. (deloitte.com)
