AI monoculture: individually accurate, collectively fragile
Interactive model of AI concentration risk: 20–50 firms with identically accurate AI models, a dominant vendor or shared foundation model, and the resulting distribution of simultaneous failures versus independence, with policy levers for vendor diversity and human challenge.
Most discussions of AI risk ask whether one model is accurate enough. Financial stability depends on a different question: when models fail, do they fail together? If many lenders, insurers or trading desks run the same foundation model, the same vendor product or models trained on the same data, their errors stop being independent, and a portfolio of individually reliable decisions can become a single point of failure. The Financial Stability Board and academic researchers have flagged exactly this concern about concentration in AI providers.
This module gives every firm the same failure probability in a stress period (5% by default) and varies only how correlated the failures are, using the latent-factor structure behind the Vasicek credit-portfolio model in a two-level form: every firm shares a market-wide factor, and firms on a dominant vendor also share a vendor factor. The expected number of failing firms is identical in every setting (1.5 of 30 at the defaults). The tail is not: with 60% of 30 firms on one vendor and a same-vendor correlation of 0.6, the probability that more than a quarter of firms fail at once rises from about 1 in 12,000 under independence to almost 4%.
The bar chart compares the exact distribution of simultaneous failures with the independent binomial benchmark, a sweep shows the systemic tail climbing with vendor market share, and a simulated grid shows failures arriving in clusters. A policy table tests three levers (capping the vendor's share at 20%, halving the extra same-vendor correlation, and independent human review that catches half of errors) against simply making the shared model twice as accurate. Both human review and the better model halve expected failures, but human review cuts the chance of more than a quarter failing to about 0.7%, half what the more accurate shared model achieves (about 1.3%). The model is stylised; the mathematics of correlated defaults is standard.
Conditional failure probability (one-factor model)
P(fail | Z) = Φ( (Φ⁻¹(p) − √ρ·Z) / √(1 − ρ) ), P(K = k) = ∫ C(N,k)·P(fail|z)^k·(1 − P(fail|z))^{N−k} φ(z) dz
Each of N firms fails with probability p; ρ is the correlation of their latent drivers and Z the common factor, with standard normal distribution function Φ and density φ. Given Z, failures are independent, so mixing the conditional binomial over Z gives the distribution of the number of simultaneous failures K. The module nests this twice: a market factor shared by all firms and a vendor factor shared by the vendor's users.
Frequently asked questions
What is algorithmic monoculture?
The situation where many decision-makers rely on the same algorithm or model, so their decisions, and their mistakes, become correlated. Even if the shared model is individually more accurate than the alternatives, the lack of diversity can make outcomes collectively worse and failures systemic.
Why does AI vendor concentration matter for financial stability?
Because diversification only works when errors are independent. If a large share of the market uses one model, a single flaw, data shift or outage hits many firms at once, raising the probability of simultaneous failures far above what the individual error rates suggest. It is the same mechanism that makes correlated defaults dangerous in credit portfolios.
How can firms reduce AI monoculture risk?
By keeping errors independent: diversifying model providers, using in-house or differently trained challenger models, maintaining independent human review of AI decisions, and testing what happens when a shared upstream model fails. Making the shared model more accurate does less for the tail than restoring diversity.
Interactive explorer
The live model with sliders, charts and real-time statistics loads on this page. If it does not appear, enable JavaScript or open https://www.beyondaverages.org/ai-monoculture in a modern browser.
Related modules
- Dice & convolution: Interactive 3D dice simulator. Roll up to five dice and watch the empirical sum histogram converge to the exact convolution probabilities: uniform to triangular to bell-shaped, the central limit theorem made visible.
- Cards & hypergeometric: Interactive card-dealing simulator. Deal hands from a 3D deck and compare empirical frequencies of aces, pairs and colors against exact hypergeometric probabilities, and see why a deck's memory matters.
- Frequency: Interactive Poisson distribution explorer for insurance claim counts and hospital admissions. Move the rate λ and watch P(X=k), the mode, and the probability of zero or extreme counts update live.
- Severity: Interactive loss severity explorer comparing Log-Normal, Pareto (heavy-tail) and Gamma distributions. Adjust parameters and see the mean, median, 95th and 99th percentile claim sizes, and the share of total cost from the largest 5% of claims, update in real time.