Plus ça change. AI and the future of economic research: From Galileo back to Socrates

For centuries, economic research has been throttled by scarcity of time, labour, and access to data and methods. This column argues that while AI helps dissolve that constraint, it can also exacerbate challenges the profession is already grappling with. And by blurring the traditional distinction between observing the world and modelling it, AI may even challenge our definition of good research.

For centuries, technological innovations have reshaped how we conduct research. The printing press, for instance, reduced the cost of accessing and disseminating knowledge (Dittmar 2011). Together with reliable mail service, it enabled the creation of the first scientific journals and communal correspondence (Gotti 2014).

In economics, successive technological innovations have increasingly pushed the discipline toward a more computational, data-driven one. Three threads run through these innovations: they reduce the cost of computation, reduce the cost of data, and progressively automate the tedious work that once consumed the bulk of a research project, leaving researcher more time for what Edison called inspiration rather than perspiration. The most visible result has been a sustained rise in research output: annual CEPR and NBER working paper volumes are up approximately eightfold since the early 1980s, while submissions to top journals have roughly doubled since 1990 (Card and DellaVigna 2013, Luttmer 2024).

AI is likely to push research further along the same path. This is, by most measures, a triumph. For centuries, research has been throttled by scarcity of time, labour, and access to data and methods. AI helps dissolve that constraint. It can also improve how research is communicated and shared and help diffusing knowledge faster which is particularly relevant for policy institutions.

The validation constraint and concentration

However, AI may also exacerbate challenges the profession is already grappling with. It may worsen the validation constraint. This problem is not new. The time required to publish in top economics journals has doubled since the 1970s, reflecting a growing tendency for journals to require extensive revisions before accepting papers (Ellison 2002). The mere fact that AI lowers the marginal cost of generating research will not necessarily reduce the time and cost of peer review. If anything, the burden on editors and referees is likely to increase in an environment where AI can generate plausible but inaccurate results.

Yet, even before AI, there were signs that the publication process was already buckling under the load. The profession has tried to manage this through the use of working papers, which allow research dissemination ahead of formal journal validation, and more recently through platforms such as Substack, personal websites, and social media, which enable even faster dissemination. These channels undoubtedly accelerate the diffusion of knowledge, but come at a cost. They risk flattening the publication landscape, i.e. raising volume without necessarily increasing insight (So 2026). And they shift the validation burden onto the reader. This validation often takes the form of reputational shortcuts: attention is focused on familiar and trusted academics, institutions, or networks readers (see also Weder di Mauro 2026). The result is a literature that is simultaneously more crowded and more unequal in who gets read and cited.

AI may also increase concentration through a different channel. Frontier models carry substantial fixed costs thereby creating powerful economies of scale. Universities and institutes with access to cutting-edge models and experienced staff may pull further ahead, leaving junior researchers and resource-constrained institutions behind.

Finally, a more subtle challenge may also be brewing. Much of the informal training that shapes junior researchers has historically happened through the grind – running regressions by hand, chasing data errors, drafting first-pass literature reviews. As AI absorbs more of that grind, it risks hollowing out this training model not only at universities but also policy institutions. Efficiency gains today could therefore translate into a competence deficit later, when today’s junior staff are expected to be tomorrow’s senior validators.

What institutions should actually do

None of this argues for slowing down AI adoption. Efficiency gains are real and documented. Moreover, the profession has repeatedly risen to earlier challenges. For example, peer reviews, as we know them today, emerged in the 1970s in response to the rapid expansion of universities, journals, and submissions. There is little reason to believe the profession cannot adapt again. AI may also help to overcome the very challenges it creates. One notable example is peer reviews where AI could help accelerate the review process – an approach currently being explored through Refine’s partnerships with the American Economic Association (AEA) and the Econometric Society.¹   

Economics could also learn from other disciplines. The publication process in economics is not only slow, it also takes roughly twice as a long as in other social sciences and even four times as long as in natural sciences (Hadavand et al. 2024). These fields have undergone similar technological transformations while maintaining leaner publication processes. In the age of AI, however, merely catching up with other sciences will not be enough, we must go beyond.

Strengthen reproducibility. When distinguishing accurate from plausible but inaccurate results grows harder, transparency becomes indispensable. Journals and research institutions should move towards stricter requirements for the publication of code, data, and methodological documentation, possibly even at the working paper stage.

Develop shared norms. Scholars already acknowledge research assistants, datasets, and software packages. AI tools should become part of the same disclosure framework, with shared norms for transparency and attribution in AI-assisted research

Rethink the publication system itself. The rise of working paper series, blogs, and platforms reflects a genuine demand for faster dissemination of ideas. Yet they are imperfect substitutes for formal validation. Rather than forcing all research through one model, the profession could benefit from a more differentiated ecosystem. Economics lacks the type of shorter, modular formats other fields offer. Medicine serves as a useful template: leading journals offer a compact peer-reviewed format under 1,500 words, scoped tightly enough to that reviewers cannot demand endless additional robustness checks. A similar approach could help manage the growing volume while keeping publications timely and subjected to meaningful vetting. It could also open new entry points for junior scholars. The challenge is namely not merely to produce knowledge more efficiently, but to ensure new talent continues to have pathways into the scientific conversation.

Game changer?

This column’s assumption so far is that AI will accelerate human-led work. In that view, AI is a powerful instrument — just as the telescope, the particle accelerator, or the computer have been. It changes the pace of science without changing its fundamental structure.

But what if that assumption is wrong? What if AI alters the shape of the process itself?

It is worth recalling how long the current process has endured. Since the scientific revolution of Bacon, Galileo, and Newton in the 16th and 17th centuries, science has relied on a remarkably stable cycle: observe, hypothesise, predict, test, and revise. Every major innovation to date has been folded into this framework. As Jean-Baptiste Alphonse Karr famously observed “plus ça change, plus c’est la même chose” (“the more things change, the more they stay the same”).

The question is whether AI could break this invariant rather than just accelerate it. The jury is still out, but there are already signs that AI may challenge our definition of good research which has stood since Galileo. According to this longstanding view, good science does more than fit observations or predict accurately. It tries to uncover the underlying structure of reality. AI may blur the traditional distinction between observing the world and modelling it. In biology, AlphaFold predicts a protein’s 3D structure from its amino acid sequence with remarkable accuracy without offering a causal explanation, solving a 50-year-old challenge in biology (Bertoline et al. 2023). But it has also opened an epistemic gap between prediction and explanation.

For centuries, scientific progress has meant building better tools for humans to understand the world. AI that improves prediction without improving our understanding raises a deeper question, one particularly pertinent for policy institutions: what happens when our most powerful tool sees patterns we no longer can? The answer may not only shape the future of science, but also the role of humans within it. We may find ourselves at a fork in the road: do we keep privileging explanations humans can grasp, accepting some loss in predictive power, or embrace opaque but highly effective models and broaden what counts as scientific knowledge, potentially quietly abandoning Popper’s falsifiability criterion?

Either way, our role may become less about having the right answers. In this respect, AI resembles the slave boy in Plato’s Meno. By asking the right questions, Socrates led him to the answers he couldn’t initially see himself. If that analogy holds, our role may become less about Galileo’s method, i.e. having the right answer, and more about recovering Socrates’ oldest skill: asking the right questions.

Socrates points to a second, more subtle human advantage AI is unlikely to overturn. In Plato’s Apology, the Delphic Oracle declares no one wiser than Socrates. Puzzled, Socrates questioned those reputed to be wise and found they all believed they knew things that they actually did not. They mistook ignorance for knowledge. His wisdom, Socrates concluded, lies in recognising what he doesn’t know. AI models possess no such humility: prompted with a question, they generate an answer regardless of whether the underlying pattern is real or meaningful. The capacity to doubt our own conclusions – to distinguish knowledge from mere output – may be the one unique but also most crucial human contribution to a scientific enterprise which is becoming increasingly populated by machines that never hesitate to answer.

Source: cepr.org