AI errors: RBI panel calls for 'tolerant supervision'

MUMBAI: An RBI panel examining the responsible use of AI in finance has urged regulators to adopt a "tolerant supervisory stance" towards mistakes made by AI systems. The idea is to allow institutions some leeway for first-time errors if they have adequate safety measures in place.

The aim, the panel argues, is to encourage innovation rather than stifle it.Such tolerance is justified, the report says, because AI is inherently probabilistic and non-deterministic. A strict liability regime that penalises every misstep could make developers overly cautious, limiting AI's ability to deliver novel solutions. This approach could be controversial as it may be seen to be shielding institutions at the expense of customers who suffer losses from AI errors.The framework rests on seven "sutras": maintain trust; keep people in control; foster purposeful innovation; ensure fairness and inclusion; uphold accountability; design for transparency; and build secure, resilient, energy-efficient systems that can detect and prevent harm.Its 26 recommendations span building better data infrastructure, creating sandboxes for AI testing, and developing indigenous models to help smaller players. Regulators are advised to draft flexible rules and apply liability proportionately. Banks are told to adopt board-approved AI policies, implement strong data governance, and safeguard customers through transparency, effective grievance systems, and robust cybersecurity.

Continuous monitoring, public reporting, and sector-wide oversight are proposed to keep AI use safe and credible.