.png)
When a handful of foundation models underpin credit decisions across hundreds of institutions, a single shared blind spot can cascade into correlated, system-wide failure. No individual bank's risk framework is designed to catch that. Monoculture concentration in AI is a systemic risk that regulators and researchers are increasingly flagging, and engineers can start addressing it before regulators intervene.
The concept is familiar from other domains. In agriculture, planting a single crop variety across vast acreage means one pathogen can wipe out the entire harvest. In software, the CrowdStrike incident of July 2024 demonstrated the same principle at industrial scale: a single faulty update crashed more than 8.5 million systems and grounded 7,000 Delta flights over five days, with total losses estimated above $5 billion.
Now apply that pattern to financial AI. When multiple banks fine-tune the same base model and call the same vendor API — often training on overlapping data as well — their risk positions are correlated, not independent. Researchers at the LSE Systemic Risk Centre have argued that AI-driven financial analytics exhibits increasing returns to scale and high entry costs, making an oligopolistic market structure likely. Today, only a few major data vendors — S&P Global, Bloomberg, LSEG — serve the bulk of the industry. Foundation model providers are consolidating along the same trajectory.
Traditional model-risk governance, anchored in the Federal Reserve's SR 11-7 guidance, focuses on whether an individual bank's model is sound. It was never designed to detect cross-institution correlation arising from shared third-party models. That gap is where systemic risk lives.
Monoculture risk operates at three levels: model weights, training data, and API endpoint. The subtlest and most dangerous is data. If the training corpus over-represents benign macroeconomic conditions, every downstream fine-tune inherits the same distributional gap. When a tail event arrives — a sudden rate shock, a regional banking crisis — models trained on that shared foundation are likely to fail in the same direction at the same time.
The feedback loops compound the problem. Banks acting on identical model outputs move markets in the same direction, reinforcing the original error. Calvano et al. have shown that independent reinforcement-learning algorithms instructed to maximise profits quickly converge on collusive pricing strategies, even without explicit coordination — a result that Danielsson et al. argue becomes structural rather than emergent when institutions share foundation models.
Adversarial fragility adds another dimension. A single prompt-injection pattern or adversarial document format that exploits a vulnerability in one foundation model exploits it everywhere that model is deployed.
Concentration would matter less if switching were easy. It is not. Proprietary tokenisers and non-exportable RLHF layers tie models to specific architectures, creating steep exit costs. Latency SLAs and rate limits bind production systems to a single provider's infrastructure. An outage at that provider becomes a sector-wide outage.
The Financial Stability Board's November 2024 report on AI and financial stability identified exactly this dynamic: heavy reliance on foundation models from few providers creates single points of failure, and the operational risk compounds when multiple institutions connect to shared AI infrastructure. Institutions that want diversity often cannot achieve it quickly enough to matter.
Engineers can start building structural diversity today, without waiting for regulation.
One approach is running parallel inference across at least two architecturally distinct models and flagging divergence above a defined threshold before any output enters a decision pipeline. The cost overhead is real but bounded and provides early warning of correlated blind spots. For high-stakes paths like credit decisioning or collateral valuation, deterministic rule-based fallbacks should remain in place, with circuit-breaker patterns that auto-revert to heuristic models when foundation-model confidence or availability drops below acceptable levels. Separately, teams should inventory which foundation models and checkpoints sit behind every critical service, track upstream training-data provenance where disclosed, and flag undisclosed provenance as a risk factor — an internal "diversity score" analogous to supplier-concentration metrics in procurement.
The EU AI Act classifies general-purpose AI models as carrying systemic risk once training compute exceeds 10^25 floating-point operations (Article 51), triggering obligations under Article 55 for risk evaluation, incident reporting, and cybersecurity measures. The OECD and FSB have both identified third-party AI dependency and correlated machine-mediated behaviour as emerging threats to financial stability. Model-diversity mandates or disclosure requirements are likely within the next few regulatory cycles.
If most institutions depend on the same foundation model, a correlated failure leaves the lender of last resort backstopping a market-wide blind spot — a scenario none of the current frameworks address. A model supply chain audit is a reasonable first step.
Interested in fast, accurate data extraction from financial statements without the hassle? Financial Statements AI has everything you need. Sign up here for a free trial.
