When every public-health system asks the same model

#Question

What happens when many public-health institutions rely on the same small set of AI models for analysis, synthesis and decision support?

#Answer in brief

They can become less independent than they appear.

A ministry, regional agency, hospital network, university and research group may each conduct its own analysis. If all of them depend on the same foundation model, retrieval provider, training lineage or evaluation assumptions, part of their error structure is shared.

The decisions may be institutionally separate but epistemically correlated.

Research on algorithmic monoculture shows that widespread reliance on the same decision algorithm can reduce the diversity of outcomes and create systemic vulnerability even when the common algorithm performs well for individual users.^[1] Recent empirical work on more than 350 large language models found substantial correlation in model errors, including among larger and more accurate systems.^[2]

For public health, that suggests a new governance question:

How many apparently independent judgments depend on the same underlying model ecosystem?

#Accuracy is an individual property; resilience is a system property

Model evaluation usually asks:

• Is the answer correct?
• Is the model calibrated?
• Does it outperform a baseline?
• Does performance generalize across groups?

Those are necessary questions.

They do not measure institutional concentration.

Imagine ten public-health organizations independently reviewing the same emerging signal.

If ten teams use ten genuinely different information pathways, their errors may partly cancel.

If ten teams query the same model with similar prompts and the same retrieval source, the outputs may converge for reasons unrelated to the strength of evidence.

The system can appear to have ten confirmations while containing one upstream judgment repeated ten times.

That is not independence.

#Monoculture does not require one vendor

The obvious version of concentration is simple:

everyone uses the same model.

But shared dependence can be deeper.

Two tools may have different interfaces while sharing:

• the same base model;
• the same training family;
• the same cloud provider;
• the same retrieval index;
• the same benchmark culture;
• the same fine-tuning corpus;
• the same external knowledge source;
• the same prompt templates;
• the same evaluator model.

Visible diversity can therefore coexist with hidden common ancestry.

The 2025 study of correlated LLM errors is important precisely because it found that error correlation can remain high even across models from different providers and architectures.^[2]

The relevant object is not vendor count.

It is dependency diversity.

#A synthetic public-health example

Suppose an unusual cluster of adverse events appears in routine data.

Five institutions independently ask AI systems to:

1. summarize the signal;
2. identify plausible explanations;
3. prioritize literature;
4. suggest whether escalation is warranted.

All five reports mention the same mechanism and downgrade two alternative explanations.

Decision-makers see convergence.

But all five systems use foundation models trained on overlapping web corpora and retrieve from similar indexed sources.

The consensus may be real.

It may also be partially manufactured by common upstream representation.

Without dependency information, the two possibilities are difficult to distinguish.

#Correlated error changes the meaning of consensus

Consensus is usually valuable because independent observers reduce the chance that one person's idiosyncratic error dominates.

That logic weakens when errors are correlated.

If every system has a 5% probability of making a particular mistake independently, agreement among several systems can be reassuring.

If the same mistake is driven by shared training data or shared model behaviour, multiple outputs do not provide equivalent independent evidence.

The number of interfaces is not the number of independent information sources.

Public-health governance therefore needs to distinguish:

multiple outputs
from
multiple independent epistemic pathways

#AI can compress methodological diversity

The risk is not limited to factual hallucination.

Foundation models can normalize:

• which study designs are suggested;
• which biases are mentioned first;
• which evidence hierarchies are privileged;
• which uncertainties are omitted;
• which statistical methods appear "standard";
• which interpretations sound professionally acceptable.

If the same model assists protocol design, code generation, evidence synthesis and interpretation across many organizations, methodological variation may narrow.

That can improve consistency when the common pattern is good.

It can also make blind spots harder to detect.

A field can become more standardized and less resilient at the same time.

#Public health already knows why redundancy matters

Surveillance systems often combine multiple data sources because no source is complete.

Laboratory data, emergency-department activity, mortality, sentinel reporting and community signals can disagree.

That disagreement is not always noise.

It can reveal which part of the system is changing.

AI governance should apply the same logic.

A second model is useful only if it contributes a meaningfully different failure mode.

A second institution is useful only if its analytical independence is not merely organizational.

#A proposed dependency map

For consequential AI-supported public-health decisions, institutions could maintain a lightweight AI dependency map.

It would record:

• base model or model family where known;
• provider;
• retrieval source;
• key external datasets;
• evaluator or judge model;
• fine-tuning or domain-adaptation source;
• human review process;
• fallback method when the AI path is unavailable.

The purpose is not to expose proprietary internals.

It is to identify common points of failure.

Two systems with different branding but identical dependencies should not automatically be counted as independent checks.

#Epistemic diversity can be designed

This article proposes epistemic diversity as a resilience property of AI-supported public-health systems.

It can be increased through combinations of:

#Method diversity

Compare model-assisted reasoning with explicit statistical, epidemiological or rule-based methods.

#Source diversity

Use different evidence repositories or independently curated evidence streams.

#Model diversity

Use meaningfully different model families when independent review is needed.

#Human diversity

Preserve reviewers with different disciplinary backgrounds and incentives.

#Temporal diversity

Reassess high-consequence questions after new evidence rather than accepting one synchronized AI-mediated interpretation.

The goal is not disagreement for its own sake.

The goal is to reduce the probability that one hidden dependency determines the entire system.

#Ensemble is not automatically diversity

An ensemble of five models can still be a monoculture.

If all five share similar training data and architectures, correlated error may remain high.

Likewise, asking one model five times does not create five independent reviewers.

The right question is:

What uncertainty becomes visible because this second pathway is genuinely different?

If the answer is none, the redundancy is cosmetic.

#Procurement becomes a public-health resilience issue

AI procurement is often treated as an IT decision.

With widespread deployment, provider concentration can become a public-health systems issue.

A procurement framework might therefore ask:

• Which decisions would fail simultaneously if this provider were unavailable?
• Which institutions depend on the same model family?
• Can critical outputs be reproduced without the provider?
• Is there an independent verification path?
• Are errors tested for correlation across systems?
• Can a model update change outputs across many organizations at once?

This is analogous to concentration risk in other infrastructures.

A highly capable service can still create systemic fragility if too much depends on it.

#Correlated error should be measured directly

Model diversity should not be inferred from logos.

For a set of systems used as independent checks, evaluate:

• disagreement rate;
• joint error rate;
• conditional error correlation;
• performance on difficult subgroups;
• sensitivity to prompt variation;
• sensitivity to source perturbation;
• shared failure on adversarial or ambiguous cases.

The ICML study on correlated LLM errors shows why this matters: systems that look different can fail together.^[2]

For public health, the relevant test set should reflect the actual decision domain.

#Human oversight does not automatically solve monoculture

A human reviewer can reduce model error.

But humans can also be anchored by the same AI output.

If every team begins from the same generated summary, human review may preserve rather than break correlation.

Meaningful human oversight sometimes requires:

• reviewing primary evidence before seeing the AI conclusion;
• independent first-pass judgments;
• explicit dissent capture;
• blind comparison of alternative analyses;
• documenting which model suggestion changed the final decision.

Human presence is not equivalent to epistemic independence.

WHO guidance on multimodal AI emphasizes accountability, transparency and appropriate human control.^[3] A resilience perspective adds another question: whether the system preserves enough independent reasoning to detect shared failure.

#A proposed concentration indicator

For a network of public-health organizations, one useful metric could be the proportion of consequential AI-supported workflows that depend on the same upstream model family or provider.

That metric alone would not define risk.

It could nevertheless reveal concentration that ordinary model-performance dashboards miss.

A stronger assessment would combine:

dependency concentration
× error correlation
× decision consequence
× availability of independent verification

This is a proposed analytical frame, not a validated risk score.

#The deeper shift

The first generation of AI governance asks:

Is this model safe enough to use?

A mature public-health system will also need to ask:

What happens if everyone uses systems that fail in the same way?

That is a different level of analysis.

It moves from model safety to system resilience.

Public health has long understood that resilience depends on redundancy, heterogeneity and the ability to detect failure.

The same principle should apply to the infrastructure used to interpret evidence.

#Practical questions

Before treating multiple AI-supported outputs as independent confirmation, ask:

• Do they share a base model or provider?
• Do they retrieve from the same information sources?
• Are their errors empirically correlated?
• Did human reviewers work independently before seeing generated conclusions?
• Could one provider update change several institutions simultaneously?
• Is there a non-AI or independently modelled verification path?
• Which conclusions would survive if the dominant model were removed?

These questions do not argue for avoiding foundation models.

They ask whether institutional confidence matches actual independence.

#Limitations and uncertainty

Algorithmic monoculture is an established concept, but most formal work has examined selection and decision systems outside public health.^[1] The correlated-error evidence cited here comes from general LLM benchmarks and a hiring task, not epidemiological decision-making.^[2] This article therefore extends those findings as a public-health systems hypothesis rather than claiming that correlated institutional failures have already been quantified in public health. The proposed concepts of epistemic diversity, dependency mapping and concentration monitoring require empirical validation. Diversity also has costs: maintaining multiple methods, providers and review pathways can reduce efficiency and complicate governance. The objective is proportionate resilience, not maximal heterogeneity.