CFTC advisers urge governance, explainability and data controls for AI in markets

Commissioner Kristen Johnson opened the Market Risk Advisory Committee’s Future of Finance subcommittee meeting by praising artificial intelligence’s potential across medicine, farming and trading while warning about the risks of deploying models without guardrails. She told the panel the commission should “increase market and credential regulators’ understanding of how the underlying technology operates and the integrity of the inputs it relies upon” and emphasized governance, explainability, data controls and testing as central priorities.

Panelists framed the issue as an extension of existing supervisory work rather than a wholesale new regulatory regime. David Felon of the CFTC said the agency is using the AI executive order and the NIST AI Risk Management Framework to guide both internal use and external supervision: “We’re reviewing how we use AI internally and how our registrants use AI,” he said, while stressing particular concern about data provenance and customer protection. The CFTC has issued a request for comment and continues supervisory exams that include IT and model changes, Felon added.

Industry and academic speakers described complementary opportunities and risks. Jason Harrell (DTCC) and other international panelists said NIST provides an overarching risk-management approach but leaves unanswered questions about how to test and validate AI systems in a sector-specific way. Chen Arod (Solidus Labs) said generative tools have enabled more scalable illicit activity — for example, bots that can adapt to evade rule-based detection — while also improving detection when applied with appropriate supervision: “When used correctly and with the right supervision… there’s been significant reductions in false positives reported by compliance teams,” he said.

Academic speakers underscored the civil‑rights dimensions of model design. Professor Kim presented research on “model multiplicity,” showing that multiple models can have comparable accuracy but distribute error differently across demographic groups. She argued that developers should search for less‑discriminatory alternatives (LDAs) before deploying models for credit, employment or housing decisions and proposed that liability and regulatory regimes recognize a developer’s duty to document such searches.

Technical experts explained why explainability and testing are hard. A panelist with a machine‑learning background described deep neural networks and how generative large language models are trained and decoded, noting that stacked layers and proprietary third‑party models can make explainability and oversight challenging for firms that lack access to training pipelines.

Throughout the discussion members raised systemic and operational concerns — vendor concentration for high‑performance AI; the difficulties of supervising third‑party models and APIs; and the potential for AI to enable more sophisticated scams (one example in Commissioner Johnson’s opening was a voice‑cloning fraud that led to a $35 million transfer). Members recommended targeted data collection or anonymized surveys to better understand current AI adoption and risk‑management practices across registrants.

The subcommittee did not adopt formal rules at the meeting. Chairs said staff will produce draft language and broad recommendations (focused on governance, explainability, bias mitigation, testing/monitoring and data controls) for subcommittee review, with at least one follow‑up meeting scheduled to shape final recommendations to the commission. The discussion will continue as the CFTC balances existing supervisory tools with principles-based guidance designed to address new AI‑specific challenges.