Research
IMF's Tobias Adrian says future flash crashes will come from many AI systems reacting to the same information at once, not from coding errors
In an IMF blog post published 23 July 2026, the Fund's Financial Counsellor argued that AI is now embedded deep enough in financial decision-making that synchronised trading behaviour, not software bugs, is becoming the primary flash-crash risk, alongside concentration risk from a handful of AI and cloud providers.
Adrian's framing draws a specific distinction worth sitting with: earlier generations of trading-related flash crashes, such as the 2010 US equity flash crash, were typically traced back to a malfunctioning algorithm or a fat-fingered order that cascaded through the market structure. The risk the IMF is describing here is different in kind, not degree. It is not one system failing, it is many independent AI systems, built by different firms, correctly processing the same news or price signal and arriving at the same trading decision simultaneously, producing a synchronised move that looks identical to a crash from the outside but has no single faulty component to blame or fix afterward.
The concentration-risk argument is the other half of the piece and gets less attention but may matter more operationally. As more financial institutions build their AI trading and risk infrastructure on a small number of cloud, data and foundation-model providers, a single outage, cyberattack or geopolitical disruption at one of those providers could simultaneously degrade risk management and trading systems across many institutions at once, a correlated operational failure layered on top of the correlated trading-behaviour risk. This is the systemic-risk vocabulary regulators used for interconnected bank balance sheets in 2008, now being applied to AI infrastructure providers that most financial institutions don't consider counterparties in the traditional sense.
The IMF's recommendations, mapping model dependencies, monitoring correlated trading strategies, folding AI risk into stress tests, are sound in principle but describe a monitoring capability that barely exists yet at most central banks and supervisors. Mapping which institutions rely on which AI providers and how correlated their trading logic actually is requires a level of firm-level model transparency that firms have strong competitive incentives to resist disclosing, and no regulator currently has the authority or tooling to compel it comprehensively. The IMF is naming the right problem well ahead of having a workable solution.
This is an argument, not a rule, and it carries no supervisory mandate on its own, but it is likely to shape how the FSB's parallel sound-practices consultation and national supervisors like the FCA and ESMA think about correlated AI risk when they move from consultation to concrete guidance later this year. Watch whether the FSB's October final report, or any follow-up IMF working paper, translates Adrian's flash-crash framing into a specific proposed metric, such as a model-concentration disclosure requirement, since that would be the first sign this analysis is heading toward enforceable practice rather than remaining a widely-cited warning.
Read the original: Business Today - IMF's Tobias Adrian says future flash crashes will come from many AI systems reacting to the same information at once, not from coding errors. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.