The Bank for International Settlements has spent two years building a case that artificial intelligence poses real risks to financial stability — and its latest work says the dangers are growing, not receding.

In its December 2025 quarterly report, the BIS grouped AI valuations alongside other fragile corners of the markets as a source of mounting instability. That shift is significant. A year ago, the June 2024 report framed AI more broadly — as an opportunity and a challenge for the economy and central banks. Now the BIS has zeroed in on something sharper: how concentrated AI investment has become, and what happens if that concentration breaks.

Concentration and the Earnings Question

The BIS found in November 2025 that AI-driven buying has clustered around established companies with real profits, not spread across speculative startups burning through venture cash. This sounds like good news — it means the AI trade isn't a classic bubble of money chasing companies with no revenue. But concentration itself is dangerous. If a handful of mega-cap names get re-rated because of earnings disappointment, a regulator's move, or a competitor breakthrough, that shock travels fast through passive index funds and other portfolios that hold the same stocks.

Tech stocks rallied in January 2025, helped partly by the arrival of DeepSeek, a Chinese AI model that challenged the industry's assumptions about how much computing power frontier AI really needs. DeepSeek's emergence unsettled the financial models propping up valuations for Nvidia and cloud giants. Prices recovered, but the BIS's underlying concern — that valuations may not be justified — never went away.

The Oversight Gap

Where the BIS sounds most urgent is on financial services. A June 2025 report from its Financial Stability Institute concluded that AI in banking and lending without proper guardrails could amplify existing financial vulnerabilities. The key point: this is not a claim that AI is inherently dangerous. It is a claim that the absence of controls is the risk. That distinction shapes how institutions should govern their use of these tools.

The channels of risk are familiar from past waves of financial technology — herding behavior, models no one fully understands, correlated failures when multiple institutions rely on the same systems — but AI operates at faster speeds and larger scale. Trading algorithms built on large language models, credit decisions made by opaque AI systems, and fraud-detection platforms that share common architecture across competing banks all create failure modes that traditional regulation was not built to catch.

The Financial Stability Board moved in the same direction. Reuters reported in October 2025 that global financial regulators had expanded their monitoring of AI-related risks. The challenge: regulatory capacity is limited, and the question is whether monitoring turns into enforceable rules before exposures pile up.

What Central Banks Are Watching

The BIS's June 2024 report flagged the long-term macroeconomic dimension: AI-driven gains in productivity could shift interest rates, change how monetary policy moves through the economy, and make inflation harder to forecast. Those are slow-moving forces. The financial stability risks in the December 2025 report operate on much faster timelines.

Central banks and supervisors are caught in a timing problem. The structural economic effects of AI will take years to show up clearly in data. But the financial system is already loaded with AI exposures — in stock portfolios, in lending models, in the operational backbone of banks that matter to the whole system — and those need oversight now. The BIS has mapped both timescales in its recent work, but the emphasis has shifted toward the faster-moving risk.

Whether existing rules and oversight tools can adapt quickly enough is an open question — and one no one can yet answer with real confidence.