Bailey calls for AI model testing standards

Bailey called rigorous model testing a sensible starting point and said it should take place both before and after deployment.[1] He argued that testing should form part of standards for deploying AI models.[1] Over time, he said, standards could promote a consistent approach across the financial system and possibly the wider economy.[1] Bailey said standards could eventually promote a consistent approach across the financial system and possibly the wider economy.[1] Bailey also said progress in AI testing must accelerate.[1]

Bailey said testing can help identify vulnerabilities and build confidence in safeguards.[1] He cautioned, however, that it would not eliminate AI failures and should not be treated as a complete solution.[1] Unexpected model behavior, in his view, is a reason testing is necessary, not proof that testing has failed.[1] He called for learning from incidents and near misses across AI development, deployment and use.[1] For teams evaluating assurance, his position frames testing as an ongoing way to find and learn about risks, rather than a guarantee of safe behavior.

Bailey said testing should not be treated as a substitute for future regulation, and that a more formal regulatory framework may emerge over time.[1] In his view, regulation was not the right place to start.[1] The distinction matters: his proposal puts testing among the foundations for AI deployment standards, while leaving open the possibility of later formal rules.[1] Bailey also emphasized that people should retain the ability to intervene in AI systems and set boundaries for them.[1] He also said people should retain the ability to intervene in AI systems and set boundaries for them.[1]

Bailey said frontier AI materially increases the scale and sophistication of cyber threats to the financial system, and that financial firms’ resilience cannot be considered separately from AI advances.[1] He said testing models used by financial firms could inform safe AI deployment in cyber defense and agentic trading and payments.[1] He also pointed to potential benefits from frontier AI, including scientific discovery, higher productivity and economic prosperity.[1] Bailey said frontier AI systems learn from human-generated information and from the products of their own reasoning.[1] He warned that a sufficiently powerful system in a self-reinforcing loop could reduce society’s ability to oversee and intervene in it.[1]