Should We Really Keep AI Out of Epidemiology?

AI can make epidemiology faster. A new paper argues that this may be exactly the problem. The warning deserves to be taken seriously, but the answer may be better automation, not less of it.

Artificial intelligence has arrived in public health at an awkward moment.

Researchers already face overwhelming literature, increasingly complex datasets, pressure to publish, limited funding, and a seemingly endless list of analytical decisions. Generative AI promises help with many of these problems. It can search and organize information, generate and debug code, summarize evidence, interrogate datasets, assist with writing, and increasingly coordinate multiple steps in a research workflow.

That sounds attractive.

Peter Tennant, Haidong Lu, and Daniel Westreich want epidemiologists to be much more cautious.

In their 2026 International Journal of Epidemiology article, “Why Epidemiology Should Not Be Automated: The Case for Slow Epidemiology,” the authors challenge the emerging vision of highly automated science. Their central concern is that scientific discovery depends upon critical reasoning about data, context, causality, and the limits of existing knowledge. Those are precisely the activities that could be weakened when researchers begin handing intellectual work to machines.

Their alternative is what they call Slow Epidemiology: fewer studies, better questions, stronger designs, more transparent reporting, deeper sensitivity analyses, and researchers who remain intellectually engaged with their work.

There is a lot here worth agreeing with.

But there is also an important question hiding underneath the argument:

Does automation necessarily make epidemiology shallower?

I don’t think it does.

And the distinction may determine whether AI ultimately makes public health science better or simply makes more of it.

The Case Against Automated Epidemiology

Tennant and colleagues begin with a straightforward concern. Efficiency is not the same thing as scientific progress.

A system that produces an analysis quickly has not necessarily produced a good analysis. Epidemiology requires investigators to understand where data came from, why variables relate to one another, what assumptions an analysis requires, which biases threaten an inference, and whether a statistically elegant answer corresponds to anything meaningful in the real world.

The authors argue that current generative AI is poorly suited to that kind of reasoning.

For example, an AI system asked to identify possible confounders can retrieve variables commonly discussed in previous studies. That is different from constructing a defensible causal model of the process generating the data. The former involves recognizing patterns in existing knowledge. The latter requires understanding why variables occupy particular causal roles.

This leads to their first major concern.

AI could make science more conventional

AI systems learn heavily from what humans have already produced, which creates an interesting problem for scientific discovery. If yesterday’s literature disproportionately studied certain populations, favored certain questions, or repeatedly relied on particular methodological approaches, an AI system trained on that literature may reproduce those same patterns. Ask AI what researchers should study next, and it may offer a sophisticated remix of what researchers have studied before. Tennant and colleagues therefore worry that AI-assisted hypothesis generation could reinforce existing priorities and biases while giving them the appearance of objectivity. They raise a similar concern about methodology, since AI could repeatedly recommend established analytical conventions, including conventions that later prove flawed. This is one of the paper’s strongest arguments. A scientific ecosystem increasingly shaped by models trained on previous science could create a feedback loop in which past science shapes AI training, AI shapes new research, and that research becomes training material for the next generation of models. Without deliberate efforts to disrupt that cycle, scientific novelty could gradually become increasingly sophisticated recombination.

The Cognitive Offloading Problem

The second concern is even more interesting. Suppose AI becomes good enough that researchers routinely rely on it, even if it remains imperfect. Researchers might ask AI to generate code, suggest variables, conduct preliminary analyses, or draft interpretations, with humans reviewing the work afterward. On its surface, this sounds like responsible human oversight. Tennant and colleagues argue that it may not work that way because humans are susceptible to automation bias. When automated systems perform most of the time reliably, people can become less attentive to the moments when those systems fail, gradually turning the supposed “human in the loop” into something closer to a human rubber stamp.

There is also a deeper concern about how scientific expertise develops. A doctoral student becomes a good epidemiologist partly by struggling with research questions, drawing bad causal diagrams, writing broken code, discovering why analyses fail, revising models, reading papers closely, and defending decisions to skeptical colleagues. Those tedious and sometimes frustrating activities are not simply obstacles standing between researchers and more interesting intellectual work. They are part of how researchers learn to do that intellectual work in the first place. Tennant and colleagues make this point forcefully: if researchers outsource question formulation, study design, analysis, and writing, they may also outsource some of the experiences through which scientific expertise develops. Here again, I think they are right.

Where I Part Ways With Slow Epidemiology

The disagreement comes from what follows. The paper moves from the compelling observation that scientific reasoning should not be surrendered to AI toward a broader argument against automating epidemiology, but those are not quite the same proposition. Consider how much of research involves tasks such as searching databases, screening abstracts, cleaning and merging datasets, modifying code, creating tables, checking documentation, converting files, tracking references, comparing model specifications, identifying inconsistencies in manuscripts, and completing reporting checklists. Collectively, these activities consume enormous amounts of scientific labor, yet they vary considerably in how much human judgment they require. Automating some of them does not necessarily remove humans from science. It can remove friction from the scientific process and create more time for the reasoning that Tennant and colleagues rightly want to protect.

The more useful question, then, is not whether epidemiology should be automated, but which parts should be automated, which should be augmented, and which should remain fundamentally human. Framing the issue this way creates a very different conversation about the role AI should play in public health research.

Automation Is Not the Same as Delegation

Imagine an epidemiologist developing a causal analysis. In one version of AI use, the researcher hands over the data and asks the system to determine the causal structure, select the appropriate covariates, run the analysis, and explain what the results mean. That is delegation, and it deserves much of the skepticism Tennant and colleagues express. But consider a different workflow. The investigator develops the causal model, while AI interrogates the decisions behind it. Why is this variable considered a confounder? Could it instead be a mediator? What unmeasured common causes might explain the observed relationship? Which assumptions must hold for the proposed identification strategy? What sensitivity analyses could challenge the conclusion? The researcher answers those questions, the AI challenges the answers, and the researcher revises the model accordingly. That is not automated epidemiology. It is AI-assisted deliberation, with the researcher retaining responsibility for the scientific reasoning while using AI to probe assumptions and expose potential weaknesses. Used this way, AI could potentially make epidemiologists think more, not less.

AI Can Be Used to Create Friction

This may sound strange because nearly every technology company sells AI with the same basic promise: faster, easier, less work. Science may need a different philosophy. What if one of AI’s most valuable uses is occasionally making research harder?

Imagine an epidemiology assistant instructed not simply to generate an analysis, but to attack one. It could ask researchers to clearly define their estimand and hypothesized data-generating process, identify the assumptions required to estimate an effect, consider competing causal structures, examine which populations are missing from the data, specify what findings might falsify their interpretation, identify analytical decisions that could materially change the conclusion, and search for evidence that contradicts their preferred explanation. A system designed this way would not replace scientific reasoning. It would scaffold reasoning, functioning as an always-available methodological skeptic that pushes investigators to articulate and defend their decisions. Unlike a tired collaborator on Friday afternoon, AI can also keep asking annoying questions indefinitely.

The Paper Identifies a Very Real Threat: Scientific Slop

Another place where Slow Epidemiology and an AI-positive position should align is here.

Science already has a quantity problem (hence, why we exist)

Tennant and colleagues point to incentives that reward researchers for producing large numbers of papers and warn that generative AI can dramatically increase the production of derivative or low-quality research. Publishers already face the possibility that worthwhile research could become increasingly difficult to find amid an expanding volume of questionable work.

Imagine reducing the effort required to produce a mediocre manuscript by 90%. The predictable outcome isn’t automatically ten times more scientific discovery. It could simply produce ten times as many mediocre manuscripts.

Literature searches become harder. Peer review becomes more burdensome. Meta-analyses ingest more questionable studies. Researchers spend more time separating meaningful signals from machine-generated noise.

We could automate ourselves into a scientific landfill.

The answer, though, isn’t necessarily slower technology. It is higher standards.

AI Should Raise the Floor and the Ceiling

This is where the article’s most important idea may actually point toward a different AI future. Tennant and colleagues argue that epidemiology should respond to automation by raising scientific standards through clearer questions, more rigorous designs, more transparent reporting, and more thorough sensitivity analyses. I agree, but I also think AI can help researchers meet and enforce those higher standards. Research systems could automatically check adherence to reporting guidelines, independently reproduce analyses and flag discrepancies, search for contradictory evidence, run sensitivity analyses across alternative specifications, or verify whether cited papers actually support the claims attributed to them. We could also assign AI systems different adversarial roles, asking them to evaluate the same study from the perspectives of a statistician, epidemiologist, subject-matter expert, skeptical reviewer, or community stakeholder. Rather than building a pipeline that moves directly from AI to paper, we could create one in which the researcher develops the work, AI critiques it, the researcher revises it, independent AI systems verify it, and humans ultimately review and remain accountable for the resulting scientific record. In that model, automation increases scrutiny alongside productivity. That represents a very different technological trajectory from simply using AI to produce more research faster.

The Bigger Risk May Be Bad Incentives

There is another reason I am reluctant to blame automation itself. AI did not create the publish-or-perish culture, journals that reward novelty over replication, incentives to divide a research program into the maximum possible number of publications, or the analytical flexibility that can produce questionable findings. Humans built those systems, and AI can certainly accelerate their worst tendencies.

The greater danger, however, is not simply that AI makes science faster. It is that we are applying faster technology to a scientific incentive system already optimized for quantity. Give that system powerful automation, and it will predictably produce more of what it has already been rewarded to produce: more studies, more analyses, and more papers. But the same amplification could work in another direction. If scientific institutions place greater value on reproducibility, methodological transparency, replication, data quality, useful null results, community relevance, and genuine contributions to cumulative knowledge, AI could help accelerate those practices as well. Technology tends to amplify the system in which it operates, which means the arrival of AI should prompt us to look closely not only at what the technology can do, but at what our scientific system currently rewards.

There Is Also an Equity Question

One additional consideration deserves attention: high-quality epidemiology is expensive. Well-funded universities can assemble teams of statisticians, programmers, information specialists, methodologists, research assistants, grant writers, and scientific editors, while small public health departments, community organizations, and researchers at less-resourced institutions often cannot.

AI can change some of the economics of expertise. A researcher at a small institution may suddenly have assistance debugging R code, a local health department may be better equipped to analyze datasets that previously sat untouched, and a community organization may gain greater ability to search and synthesize a scientific literature that once required specialized staff and infrastructure. None of these tools magically turns someone into an epidemiologist, nor do they eliminate the need for methodological expertise. But expanding access to scientific capabilities is not trivial from a public health perspective. A philosophy of Slow Epidemiology should therefore be careful that “slow” does not unintentionally become science for organizations wealthy enough to afford slowness. The question is not simply whether AI introduces new risks into epidemiology. We should also ask who gains scientific capacity when sophisticated research tools become dramatically cheaper and more accessible.

So Should Epidemiology Be Automated?

So, should epidemiology be automated? Not completely. On that point, Tennant and colleagues and I probably agree more than the title of their article initially suggests. Epidemiology is not an assembly line, and its most consequential decisions involve causal reasoning, theory, ethics, context, measurement, interpretation, and scientific judgment. Researchers should remain responsible and accountable for those decisions.

But protecting human judgment does not require preserving every existing research task simply because humans currently perform it. A more productive principle for AI-era epidemiology is to automate the tedious, augment the analytical, interrogate the consequential, and keep humans accountable for the scientific claims that emerge. That approach preserves the reasoning at the heart of epidemiology while leveraging tools that can make the surrounding scientific process more efficient, rigorous, and accessible.

Maybe We Need Slow Thinking and Fast Tools

Slow Epidemiology is deeply appealing. Tennant and colleagues remind us that good science requires time, noting that a rigorous target trial emulation can take 12 to 24 months and arguing that fewer excellent studies may ultimately contribute more to population health than an avalanche of derivative research. That lesson should not disappear simply because our tools become faster. But faster tools and slower thinking are not opposites. They may be complements. Machines can search faster, process information faster, examine more possibilities, reproduce analyses, identify inconsistencies, and challenge our assumptions. Researchers can then use some of the time and cognitive capacity those tools recover to think more carefully about what the numbers actually mean. The future worth building is not one-click epidemiology. It is a form of epidemiology in which researchers have extraordinarily powerful machines at their disposal while remaining deeply responsible for the questions they ask, the assumptions they make, and the conclusions they draw. In that sense, perhaps the goal is not to choose between Slow Epidemiology and AI at all. It is to combine slow thinking with fast tools.

Slow thinking. Fast tools. Better science.

That is an AI future that public health should be willing to explore.