AI can assist the workflow; it should not own the epidemiological decision

#Question

Which decisions should remain explicitly human when AI assists epidemiological work?

#Answer in brief

The useful boundary is not that AI may write text while humans make decisions. Epidemiological work contains many decisions, some trivial and some constitutive of the scientific question.

Human ownership should remain explicit where a choice changes what population is studied, what exposure or intervention means, which records are included, which causal assumptions are accepted, whether a model result is credible, how uncertainty is interpreted, whether a signal triggers action, or what claim is published.

WHO's general AI-for-health guidance places protection of human autonomy, transparency, responsibility and accountability among its core principles.^[1] Its 2025 guidance on large multimodal models specifically addresses use in scientific research and public health.^[2]

AI can assist those decisions. It should not silently own them.

#Not every human click is meaningful oversight

A common governance fiction is that a human clicked approve, therefore the workflow is human-controlled.

That is weak oversight if the reviewer cannot inspect the inputs, does not understand how the output was generated, has no time to challenge it, has no authority to reject it, or is systematically shown only one option.

Meaningful human control requires decision rights and usable evidence.

#Classify AI assistance by consequence

The three-class scheme below is a proposed editorial-methodological framework for this article, not a WHO or OECD taxonomy. It operationalizes the broader principles of human autonomy, accountability and traceability into research-workflow decisions.^[1][3]

A practical workflow can separate three classes.

Low-consequence support includes bibliography formatting, terminology suggestions, table conversion and code comments. These tasks can often be automated with lightweight review.

Evidence-sensitive support includes candidate source discovery, data-cleaning suggestions, code generation, extraction of study characteristics and statistical model suggestions. These tasks require verification against authoritative sources, data and executable results.

Decision-defining support includes choosing eligibility, deciding whether a variable is a confounder, accepting an imputation strategy, deciding a surveillance signal is actionable, interpreting a causal estimate and authorizing publication.

These decisions require named human responsibility.

The categories are about consequences, not about how many tokens were generated.

#Question definition must remain accountable

An AI system can propose research questions.

But choosing the question sets the scientific object.

Compare a broad question about whether vaccination is associated with hospitalization with a target question specifying the eligible population, vaccination strategy, comparison strategy, time zero, follow-up and outcome.

If an AI silently narrows or broadens those components, downstream analysis can be internally consistent while answering a different question.

A human investigator should approve the final target question and its rationale.

#Eligibility and exclusions require ownership

AI can identify records that look anomalous.

The decision to exclude them can change the study population.

A safe workflow records the AI suggestion, reason, records affected, human decision and implemented rule.

The reviewer should be able to reject the suggestion or request an alternative.

AI-assisted cleaning should never mean undocumented removal.

#Causal assumptions cannot be delegated to correlation discovery

A model may detect that a variable predicts both exposure and outcome.

That does not establish that it is a confounder.

Causal roles depend on substantive structure and timing.

An AI can help draft a DAG, identify plausible pathways, retrieve supporting literature and challenge missing assumptions.

The human investigator remains responsible for the causal model used to justify adjustment and interpretation.

#Code generation needs execution and review

AI-generated code can accelerate routine transformations and analysis.

The output should still pass through tests, data checks, inspection of affected records or summaries, sensitivity analysis where consequential, and code review for material analytical steps.

The important distinction is between generated code and accepted analytical method.

Execution success proves only that the code ran.

#Model acceptance is a scientific decision

An AI assistant might recommend a generalized linear model, spline, random effect, transformation or threshold.

The decision should be tied to the estimand, data-generating process, diagnostics, assumptions and alternatives.

Otherwise, model recommendation becomes hidden authority.

#Interpretation is a high-risk compression step

A table with confidence intervals contains more nuance than a one-sentence conclusion.

AI is particularly useful at compressing information into prose.

That strength creates risk.

Associated with can become reduced. Compatible with no effect can become no effect. A subgroup estimate with high uncertainty can become works only in subgroup X.

The human author should approve the final scientific claim against the actual result and design.

#Escalation should be designed

Some outputs should trigger mandatory human review.

Examples include a result contradicting the prespecified analysis, a data-quality rule removing an unexpectedly large fraction of records, an AI recommendation changing the primary model, a citation that cannot be verified, a surveillance alert that would trigger costly action, or an interpretation crossing from association to causation.

This is more concrete than a generic human-in-the-loop statement.

It defines when the loop must stop automation.

#Accountability should be attributable

OECD's current AI Principles state that AI actors should be accountable according to their roles and should ensure traceability of datasets, processes and decisions across the AI lifecycle.^[3]

In research, the analogous question is: who can explain and defend this decision?

If the answer is only that the model suggested it, responsibility is missing.

A useful record names the decision owner, evidence reviewed, alternatives, version and final disposition.

#A synthetic AI-assisted analysis

An AI assistant reviews a routine-health dataset and proposes excluding everyone with more than five encounters in a month as likely duplicate utilization.

The suggestion removes 8% of records.

Human review shows that many high-frequency users are genuine patients with severe chronic disease.

If the rule had been accepted automatically, the study population would change, disease severity would be selectively reduced, utilization outcomes would be biased, and the workflow could still be completely reproducible.

Human review does not guarantee the correct decision. It creates an accountable checkpoint where domain knowledge can challenge automation.

#Where AI can add real value

The point is not to constrain AI to trivial work.

Useful assistance can include searching large source spaces, generating candidate analyses, checking code, surfacing inconsistent definitions, proposing sensitivity analyses, comparing interpretations and generating structured provenance records.

The stronger the assistance, the more important explicit decision boundaries become.

#A minimal decision-rights table

The following division is likewise a proposed workflow rule, not a quoted external standard. A useful division is: AI can propose question alternatives, while humans own the final estimand; AI can flag data anomalies, while humans own material exclusions; AI can draft causal pathways, while humans own accepted assumptions; AI can generate analytical code, while humans own the primary analytical specification; AI can draft summaries, while humans own the scientific claim and release decision.

Different regulated contexts may require stricter controls.

#Source verification is a decision boundary

AI systems can retrieve or suggest literature quickly.

But a source recommendation is not evidence until the source is actually inspected.

For material claims, human review should verify:

• that the source exists;
• that bibliographic details are correct;
• that the relevant passage supports the claim;
• that the population and setting are applicable;
• that newer guidance has not superseded it.

This boundary matters because fluent summaries can make a weak or inapplicable source feel stronger than it is.

#Uncertainty should be allowed to survive automation

AI systems are often optimized to produce an answer.

Scientific work sometimes requires a different output:

The available evidence is insufficient to decide.

A safe workflow therefore needs a valid unresolved state.

If every step must end in a recommendation, automation can convert uncertainty into false closure.

Human decision rights include the authority to hold a claim, request more evidence or stop a publication pathway.

#Oversight should be tested, not merely declared

A team can evaluate whether human oversight is meaningful by running challenge cases.

For example:

• inject an obviously wrong citation and test whether review catches it;
• propose a post-exposure covariate as a confounder and see whether the workflow challenges it;
• make an exclusion rule remove a large subgroup and verify that escalation occurs;
• give conflicting authoritative sources and test whether the system preserves disagreement.

These exercises test the governance mechanism itself.

A human-in-the-loop label without such evidence can become ceremonial. This challenge-testing suggestion is the author's proposed governance practice; the cited WHO/OECD materials support accountability and oversight principles, not this exact test protocol.

#Limitations and uncertainty

WHO's AI guidance is normative and governance-oriented; it does not prescribe one universal epidemiological workflow.^[1][2] The decision-rights framework here is an editorial-methodological proposal, not a regulatory standard. Appropriate oversight depends on task risk, domain, local law, available expertise and system capability. Human review can itself be biased, inattentive or wrong; accountability is necessary but not sufficient for validity.