Andrew Bailey calls for rigorous AI testing before formal regulation

Bank of England Governor Andrew Bailey said in an article published in the UK on 30 September that rigorous testing of advanced AI, before and after deployment, should come before the design of formal regulation. He argued that testing must reveal vulnerabilities and show whether safeguards allow people to intervene, a question he said matters for the stability of the financial system.

Bailey called AI risks ‘real and increasingly significant’ but said policymakers should first identify the problem they are trying to solve. ‘Regulation is not, in my view, the right place to start,’ he wrote. His article sets out an approach to governance; it does not announce a new rule or a timetable for one.

What Andrew Bailey wants AI testing to establish

Bailey proposed rigorous tests both before an AI model is deployed and after it enters use. Those tests, he said, should help people understand how increasingly complex systems behave, identify vulnerabilities and assess confidence in the safeguards intended to contain them. He also wants the results to inform standards for deploying models.

His central question is whether society can retain the ability to intervene as frontier AI grows more capable. Bailey described these systems as able to draw on a vast stock of human knowledge and, increasingly, to use their own outputs to improve their performance. Without effective intervention, he warned, that process could become a self-reinforcing loop that weakens meaningful human oversight.

Testing has limits, Bailey acknowledged. Models may behave unexpectedly, and tests will not eliminate failures. He said incidents and near misses during development and use should also inform safeguards. In his account, an unexpected result is a reason to learn more about a system’s behaviour, not proof that testing has no value.

Why the Bank of England governor raised financial stability

Bailey linked his argument to cyber threats facing finance. Frontier AI, he wrote, materially increases their scale and sophistication. Payments networks, financial market infrastructure, banks and other critical institutions cannot assess their resilience separately from advances in AI, particularly as firms bring the technology into their own operations.

He said testing the models used by financial firms could help establish how to deploy them safely in cyber defence, agentic trading and payments. That knowledge might eventually support consistent standards across the financial system and perhaps the wider economy. Bailey presented those standards as a possible outcome of further work, not as requirements already adopted.

The Financial Conduct Authority has separately reported that frontier AI can help firms find and analyse cyber vulnerabilities faster, while also potentially amplifying threats to firms, customers, market integrity and financial stability. Its review described firms’ experiences with the technology and expressly said it created no new rules, guidance or regulatory expectations.

Firms told the FCA that faster vulnerability discovery can outpace their capacity to check and fix the findings. The regulator said useful AI deployment depends on the tools and controls around a model, clear responsibility for decisions and human expertise to judge which weaknesses are credible and urgent. More findings do not automatically mean a firm can act on them faster.

What remains to be worked out

Bailey pointed to work by the UK AI Security Institute, which says it researches advanced AI capabilities and impacts and develops and tests ways to reduce risks. He said progress on testing must accelerate as model capabilities advance, especially when systems are used beyond controlled test environments.

His article does not set out a detailed testing regime, identify who would set common standards or specify how intervention would work across different systems. Bailey said a more formal regulatory framework may emerge over time. For now, his stated sequence is to understand model behaviour, test safeguards and establish credible ways for people to intervene.