AI, Comparative Advantage, and the Next Decade of Ecological Research

Abstract

Science is a highly, but not wholly, competitive enterprise. Scientists give considerable weight to priority of discovery. Watson and Crick raced Pauling to the structure of DNA; Darwin hurried the Origin into print after Wallace's letter arrived from the Malay Archipelago. Yet science also depends on cooperation in ways that markets do not. Researchers share methods, publish negative results (at least in principle), and build openly on each other's work. This tension between competition and cooperation is part of what keeps the enterprise moving forward. It also means that the strategic landscape of research (who has an advantage, and why) shifts as tools and norms change.

I want to argue that we are very likely approaching two such shifts in ecology, and that the second one will surprise many people who are focused on the first.

Economists since Ricardo have understood that trade benefits both parties when each specializes in what they do at lowest relative cost. The insight applies to research teams as much as to nations. A lab with deep expertise in Bayesian hierarchical modelling and a lab with a 20-year demographic dataset on desert rodents are each positioned to do work that the other cannot easily replicate. Their collaboration is productive precisely because their advantages are complementary.

The science-of-science literature has increasingly recognized this logic, even when it does not use the Ricardian label. Xiang et al. (2025) reviewed the growing literature on research topic choice, emphasizing how researchers select topics, methods and collaborators given their skills, networks, and institutional settings. Tripodi et al. (2020) showed that scientists' research trajectories are shaped by knowledge proximity and social relatedness: moving into a new area is easier when prior expertise and existing collaborations reduce the learning cost. That is a comparative advantage mechanism, even if the language is about networks and friction rather than trade theory. Scientists' rate of topic switching is increasing (Zeng et al. 2019) and Venturini et al. (2024) found that collaboration facilitates these switches, especially collaboration with productive, high-impact researchers. That is, comparative advantage in science is often co-produced by the people around you.

Topic choice has direct consequences for who gets funded. Hoppe et al. (2019), analysing NIH applications, found that topic choice alone explained more than 20% of the funding gap between Black and white scientists after controlling for other factors. What you choose to work on interacts with the reward structure of the field.

At any given moment, then, the distribution of comparative advantages across the research community shapes what science gets done. When a particular skill is scarce and productive, those who possess it enjoy disproportionate influence. When that skill becomes widespread, the advantage evaporates and the landscape rearranges.

Right now, it seems to me, quantitative and computational fluency confer a strong comparative advantage in ecology. Researchers who can build simulation models, fit complex statistical frameworks to large datasets or deploy machine learning for image classification and species identification have become central in a field organized around synthesis. These techniques require expertise. They are unevenly distributed across the community. And they are highly productive. A skilled computational ecologist can contribute meaningfully to dozens of collaborations simultaneously because the bottleneck is rarely the data but frequently the analysis.

This advantage is beginning to erode. AI-assisted research tools now perform literature synthesis, generate and debug code and automate substantial portions of the analytical pipeline. Lu et al. (2026) recently demonstrated a system called ‘The AI Scientist’ that autonomously generates research ideas, designs experiments, runs analyses and writes manuscripts. One of its papers passed peer review at a machine learning workshop. The system has obvious limitations. But the trajectory is that analytical tasks that currently require years of training are becoming accessible to anyone with a laptop and a subscription.

This is the first horizon. Within a few years, AI fluency will be table stakes for ecological research rather than a distinguishing skill. The comparative advantage of the computational ecologist will flatten as the tools democratize. In the language of Tripodi et al. (2020), the knowledge proximity required to enter computational territory is collapsing. The learning cost that once protected the specialist is being eliminated by the tools themselves.

If everyone can do sophisticated analysis, what becomes the scarce resource?

Not ideas. When Sutherland et al. (2013) convened 388 ecologists and asked them to identify fundamental research questions, they received 754 submissions, which were distilled to 100 priorities spanning every major subfield. Kuebbing et al. (2018), surveying 1179 ecologists, found the same pattern: respondents identified a wide range of important questions but cited short-term funding mechanisms and logistical barriers as the primary obstacles to pursuing them. The bottleneck in ecology has rarely been a shortage of questions worth asking. It has been the capacity to answer them.

Some might argue that the advantage will shift to theorists, to those who generate the most original concepts or frameworks. But Stern (2004) has shown that scientists accept substantial wage penalties (around 25%) to work in environments that allow them to pursue their own research agendas, and Roach and Sauermann (2010) found that PhD students self-select into academic or industrial careers based on heterogeneous preferences for independence, publishing and basic research. These results suggest that intrinsic motivation and self-selection already sort researchers toward idea-rich environments. The constraint on progress is not the supply of questions but the resources to pursue them.

The scarce resource will be original data. To see why, consider Donoho's (2024) concept of frictionless research exchange. Donoho describes a mode of computation-driven research in which code, data and methods are shared seamlessly—an ecosystem where researchers build on each other's resources with minimal friction. In such an ecosystem, any published dataset can be reanalyzed, recombined and extended by anyone. Curated datasets lose their exclusivity. The analytical pipeline is reproducible by design.

What cannot be frictionlessly exchanged is the capacity to generate new data: a 30-year census of forest plots in Panama, a network of experimental ponds manipulated across a gradient of biodiversity, access to a BSL-4 laboratory, long-term relationships with communities that permit fieldwork on their land, specialized equipment such as custom-built biometric monitors, environmental DNA sampling platforms or drone-based remote sensing rigs that not every lab can afford, build, or operate.

These assets are irreproducible in the way that a clever R script is not. When analytical capability is abundant and freely shared, comparative advantage will shift to those who control the means of data production: the experimentalists, the field ecologists and the researchers embedded in long-term monitoring programs.

This is the second horizon. It does not require that AI tools fail or stall. But requires only that they succeed so thoroughly that computation ceases to be a distinguishing capability. To me, that seems very likely.

None of this means that quantitative training becomes unimportant. A researcher who cannot critically evaluate an AI-generated analysis will be at a disadvantage regardless of their data. But the strategic calculus for departments, funding agencies and individual researchers is about to shift.

Abramo et al. (2024) recently asked whether research universities specialize in disciplines where they hold a competitive advantage. With one exception among 63 Italian universities, the answer was no: institutions generally do not concentrate resources in the fields where they perform best. The same question will soon apply at every level. Departments that invested heavily in computational hires may find those skills commoditized. Funding agencies that prioritized development of analytical methods may rebalance toward infrastructure that supports long-term data collection.

For the past two decades, the message to young ecologists has been consistent: learn to code, learn statistics, learn machine learning, learn ‘modelling’ (GIS, differential equations, species distribution modelling, etc.). That message was correct and remains so. But it is incomplete. The researchers who will hold a durable comparative advantage in the decades ahead may be those who invested in the patient, expensive, logistically difficult work of building field programs, maintaining experimental infrastructure and cultivating the expertise required to collect data that no one else can collect.

The first horizon is arriving now. The second is not far behind.

The author takes full responsibility for this article.

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.