Your industry is converging on two or three foundation models. That convergence is efficient, fast, and cost-effective. It is also creating a category of correlated systemic risk that no existing standard names, measures, or mitigates. This post introduces the framework.
Imagine every major bank, insurer, and retailer in a market using the same credit scoring model from one vendor. Regulators banned that arrangement decades ago because correlated model failure, one model, one blindspot, one adversarial pattern, produces correlated errors across an entire industry simultaneously. Systemic, not local.
Foundation models have recreated this dynamic at a scale those regulations never anticipated. A handful of model families now underpin fraud detection, content moderation, contract review, clinical triage, and supply chain forecasting across thousands of organizations. When a model update introduces a new behavioral drift, when a novel adversarial prompt bypasses a shared safety layer, or when a regulatory authority rules against a specific model's outputs, every organization that depends on that model makes the wrong call at the same moment.
This is not a vendor concentration problem. It is a cognitive concentration problem, and it has no name in current risk taxonomy. Board risk committees, NIST AI RMF, ISO/IEC 42001, and the EU AI Act all address individual model risk. None address what happens when the industry's cognitive substrate converges.
This post names the problem, introduces two tools for measuring and managing it, and gives CROs and CISOs a concrete action framework before regulators force the conversation.
Cognitive Monoculture Risk (CMR) is the systemic exposure that emerges when a critical mass of organizations within a sector or market relies on a shared cognitive substrate, specifically the same foundation model family, training corpus, or RLHF alignment procedure, for consequential decisions. CMR manifests not as individual model failure but as correlated, sector-wide decision errors triggered by a single shared blindspot. The term originates with this work and is subject to the licensing notice in the footer.
CMR is frequently misclassified as a vendor concentration risk or a supply chain dependency, categories that already exist in enterprise risk frameworks. The distinction matters because the mitigations are entirely different.
Vendor lock-in describes a procurement dependency: if a vendor raises prices, exits a market, or discontinues a product, organizations face switching costs. The risk is operational and financial. The mitigation is contractual: multi-vendor agreements, data portability clauses, interoperability standards.
CMR is a cognitive dependency: if a model develops a systematic blindspot, all organizations using that model acquire the same blindspot simultaneously, regardless of which vendor contract they signed. You can run the same base model on Azure, AWS, and your own infrastructure and still have 100% CMR concentration. You can have three vendor contracts and zero cognitive diversity.
Vendor concentration is measured in procurement contracts. Cognitive Monoculture Risk is measured in the proportion of consequential decisions routed through models sharing the same training provenance, alignment procedure, or capability boundary. These are orthogonal dimensions, and existing frameworks address only the first.
The second distinction is the failure mode. Vendor lock-in failures are sequential and detectable: one organization fails to migrate, then another. CMR failures are simultaneous: when an adversarial prompt bypasses GPT-4's content policy, every organization routing content moderation through GPT-4 is exposed in the same moment. The failure is not amplified by concentration, it is made simultaneous by it.
Foundation models acquire blindspots through four pathways, each of which transfers across every deployment of the model:
Training corpus gaps. A model trained predominantly on English-language text from certain domains will systematically underperform on out-of-distribution inputs across every deployment. Every organization using that model inherits the same out-of-distribution boundary, regardless of their fine-tuning or prompt engineering.
RLHF alignment artifacts. Reinforcement learning from human feedback introduces systematic preference biases. Research on constitutional AI and preference optimization has shown that alignment procedures can introduce consistent output patterns not intended by deployers [4, 5]. These patterns are shared across all deployments of the aligned model.
Capability ceilings. Every foundation model has a benchmark profile: tasks it performs reliably, tasks where it is unreliable, and tasks where it confidently produces incorrect output. Every organization that deploys the model inherits the same ceiling. When a market-wide decision type (e.g., evaluating a new financial instrument structure, assessing a novel supply chain risk) sits near a model's capability ceiling, correlated errors follow.
Post-deployment behavioral drift. Model updates, even those presented as improvements, can shift behavior in consequential dimensions. A financial institution relying on model outputs for credit decisions may experience different error rates after a model update, without changing any of their own infrastructure. If the update affects all deployments simultaneously, correlated risk events follow.
CROs need a metric. Qualitative descriptions of model concentration do not survive a board risk committee without a number. The following framework introduces a quantitative index designed to be computed from an organization's own AI inventory.
Model Concentration Index (MCI) is a scalar measure of an organization's cognitive concentration across its AI-powered decision workflows. For an organization with N consequential AI decision points, where each point i routes through a model family f(i) drawn from a set F of available families, the MCI is computed as: MCI = sum over all families k in F of (s_k squared), where s_k is the fraction of consequential decision volume routed through family k. MCI ranges from 1/|F| (maximum diversity, equal distribution across all families) to 1.0 (complete monoculture, all decisions through one family). The term and its index formula originate with this work.
This is structurally identical to the Herfindahl-Hirschman Index (HHI) used by antitrust regulators to measure market concentration [6], adapted to cognitive substrate rather than market share. Regulators already understand this framing: MCI above 0.75 corresponds to the "highly concentrated" threshold HHI defines, and the same intuition applies. When more than 75% of consequential decisions flow through one model family, the organization is in high-CMR territory regardless of what vendor contract it signed.
A sector-level MCI can be estimated from industry surveys on model adoption. The implication for regulators is direct: a sector with high aggregate MCI has amplified CMR even if no individual organization in that sector is in breach of any AI governance standard. The risk emerges from the composition of individual choices, not from any single choice being wrong.
Not every model concentration produces equal risk. CMR severity is a function of two dimensions: the concentration of cognitive substrate and the criticality of the decision domain.
Organizations in the upper-right quadrant, high concentration and high criticality, face the most severe CMR exposure. A bank routing credit decisions, fraud detection, and regulatory filings through the same model family is in this quadrant. An organization in the lower-left, using multiple model families for low-stakes content generation, faces negligible CMR.
The matrix also reveals where targeted intervention produces the most risk reduction: moving a high-criticality use case from a high-concentration model family to a differentiated one (even if less capable on benchmarks) may reduce systemic risk more than improving the model's average-case performance.
The mitigation for CMR is not "use fewer AI models." It is deliberate cognitive diversity: intentionally routing different decision classes through model families with differentiated training provenance, capability profiles, and alignment procedures.
This is harder than it sounds because capability differentials are real. If Model A substantially outperforms Model B on a task, a rational organization routes that task to Model A. CMR mitigation may require accepting a capability cost in exchange for systemic risk reduction, a trade-off that needs to be explicit at the board level, not left to engineering teams.
Not all AI decisions carry equal CMR weight. The following framework classifies decision domains by their correlated-failure consequence and their current model concentration, giving CROs a prioritization lens.
| Decision Domain | Typical MCI | Correlated Failure Consequence | CMR Priority |
|---|---|---|---|
| Credit / Loan Decisioning | High (>0.70) | Sector-wide credit mispricing; regulatory action | Critical |
| Fraud Detection | High (>0.65) | Simultaneous bypass by shared adversarial pattern | Critical |
| Clinical Triage / Diagnosis Support | Moderate (0.45-0.65) | Correlated missed diagnoses in same patient populations | High |
| Contract Review / Legal Reasoning | Moderate-high (0.55-0.75) | Shared clause-blindness propagating across industry contracts | High |
| Content Moderation | Very high (>0.80) | Simultaneous bypass event exposes all platforms at once | Critical |
| Supply Chain Risk Assessment | Low-moderate (<0.50) | Correlated supply chain misjudgments at macro events | Moderate |
| Marketing / Content Generation | Very high (>0.85) | Style homogenization; low consequence individually | Low |
The highest-MCI domains are not always the highest CMR-priority domains. Content generation has very high model concentration but low correlated-failure consequence. Fraud detection has slightly lower concentration but catastrophic correlated-failure consequence. MCI and CMR priority are related but not identical, which is why the MCI alone is insufficient, the decision-criticality dimension is required.
CMR management is not a technology problem. It requires a governance function that does not exist in most organizations today. The minimum viable structure for an organization with more than 10 consequential AI decision points:
Pilot governance (months 1-3): 1 Chief Risk Officer or VP Risk as executive sponsor (owns CMR as a named risk in the enterprise risk register); 1 AI Risk Analyst (owns AI inventory and MCI calculation, builds the model provenance registry); 1 Senior ML Engineer part-time (owns technical verification of model family differentiation, not just vendor names); 1 Legal/Compliance Officer part-time (maps CMR to regulatory exposure, specifically EU AI Act Article 9 systemic risk provisions [7]).
Scale-up (months 4-12): Add a dedicated AI Governance Lead; extend the model provenance registry to all business units; establish a quarterly CMR Review Board with CISO and CTO participation. Organizations subject to DORA (EU Digital Operational Resilience Act) should note that CMR maps directly to ICT concentration risk provisions that apply to financial entities from August 23, 2026 [8].
Build AI decision inventory. Map every consequential AI-assisted decision to its model family. Compute baseline MCI. Identify domains above 0.70 MCI threshold. Go/no-go gate: MCI calculated and reviewed by board risk committee.
Design routing layer for high-CMR domains. Identify alternative model families with differentiated training provenance. Pilot cognitive diversity on one critical decision domain. Gate: demonstrated MCI reduction below 0.60 in pilot domain.
Extend routing architecture across all critical domains. Establish quarterly MCI review cadence. Publish CMR as named metric in annual AI governance report. Gate: sector-level MCI below 0.65 for all critical decision domains.
Build: The MCI calculation and model provenance registry. These are specific to your AI deployment topology and cannot be purchased off the shelf. A spreadsheet suffices initially; a structured data layer is required at scale.
Buy: Model evaluation infrastructure that enables apples-to-apples comparison across model families on your specific tasks and data. Also model monitoring platforms that can surface behavioral drift as updates propagate.
Configure: API routing layer (most organizations already have one) extended with CMR-aware routing logic. This is configuration, not net-new engineering, once the provenance registry and routing rules are defined.
1. Inventory incompleteness. AI deployments exist in business units without central visibility. Shadow AI undermines any MCI calculation. Signal: unexplained gaps in the AI inventory relative to business unit headcount and AI spending. Mitigation: tie AI vendor procurement approval to inventory registration.
2. Vendor label vs. model family confusion. "We use three AI vendors" is not the same as "we have MCI below 0.5." Multiple vendors may all be reselling the same base model family. Signal: vendor contracts that mention fine-tuning of a named base model. Mitigation: require model provenance disclosure in vendor contracts, specifically base model family and RLHF procedure.
3. Capability regression on diversity. Alternative model families may underperform the primary on specific tasks. CMR mitigation may introduce task-level quality reduction. Signal: accuracy degradation in A/B testing after routing changes. Mitigation: accept quality cost explicitly at risk committee level; document the trade-off.
4. Regulatory anticipation failure. Financial regulators (EBA, FRB, PRA) are beginning to ask about AI model concentration in supervisory reviews. Organizations that have not computed an MCI will face these questions without an answer. Signal: regulatory inquiry or supervisory letter referencing AI concentration. Mitigation: compute MCI before the question arrives.
5. Update synchronization. Model providers push updates asynchronously, but in practice, major updates propagate across all customers within a short window. A behavioral change that affects all deployments within 48 hours creates a narrow but real correlated-risk window. Signal: coordinated changes in model output patterns across systems after a provider release. Mitigation: staged rollout policy with version pinning for critical decision domains.
A multinational bank has 14 AI-powered decision points across credit, fraud, and regulatory reporting. All 14 route through variants of the same foundation model family (different fine-tunes, different vendors, same base). MCI = 0.91. The risk: a single adversarial prompt pattern targeting the base model's content policy bypass propagates to all 14 decision points within the update cycle. The CRO's action: require model provenance disclosure from all three AI vendors, compute baseline MCI, and present to the risk committee with a 12-month roadmap to MCI below 0.60 for all critical domains. Regulatory framing: ICT concentration risk under DORA, which applies from August 23, 2026 [8].
A health system uses AI-assisted triage across 8 emergency departments. The triage model is the same family across all 8 sites, deployed by a single health IT vendor. MCI = 1.0 at the system level. The risk: a capability ceiling in the model's reasoning about atypical presentations produces correlated missed diagnoses across all sites simultaneously, triggered by a novel patient presentation type not well-represented in training data. The CISO's action: require the health IT vendor to disclose base model and training provenance; evaluate a second-family model for one site as a cognitive diversity pilot. Frame to the board as patient safety risk with systemic properties, not a software quality issue.
An insurance carrier routes all claims language processing and fraud scoring through two AI systems, both of which use the same foundation model family at their core. MCI = 0.88. The risk: the model family has a documented systematic bias on certain claim narrative structures (not in the carrier's training data, introduced by the provider's RLHF alignment) that causes it to flag legitimate claims from a specific demographic at higher rates. The failure is correlated across all claims processed by the model family industry-wide. The Chief Actuary's action: treat model alignment provenance as an actuarial input; require vendors to document RLHF procedures and disclose alignment bias audits as a condition of contract renewal.
A simultaneous adversarial bypass across a sector's fraud detection systems produces losses that are not independent events. They are correlated. Insurance products that cover individual fraud events may not cover correlated sector-wide events triggered by a shared model vulnerability.
Regulators asking about AI model concentration in supervisory reviews will receive either a computed MCI and a roadmap or silence. Silence is the more expensive response. EU AI Act Article 9 requires systemic risk consideration for high-risk AI; CMR is squarely within scope [7].
Computing the MCI requires an AI inventory (often already partially exists) and a provenance registry (new, but tractable). The routing architecture for diversity is configuration, not new engineering. The total cost of awareness is substantially lower than the cost of a correlated failure event.
Organizations that compute and publish MCI as part of their AI governance reporting establish a credibility signal that regulators and institutional counterparties will increasingly use to differentiate mature AI governance from checkbox compliance.
The trajectory is visible. Financial services regulators already ask about third-party ICT concentration under DORA [8]. The Bank of England's 2023 AI and ML survey asked about model vendor concentration explicitly [9]. The EU AI Act's Article 9 risk management requirements create a surface for systemic risk inquiry [7].
The logical extension is a sector-level MCI disclosure requirement: the same way banks must disclose their concentration exposure to individual counterparties, they will be required to disclose their concentration exposure to individual AI model families. Organizations that have been computing MCI internally will produce that disclosure in hours. Organizations that have not will face a regulatory gap with no fast path to closure.
The window to build this governance infrastructure proactively is open now. It will not stay open once a sector-level correlated failure event makes the risk legible to legislators who are currently not looking for it.