A year ago I argued that the introduction of artificial intelligence (AI) into economic research was not merely a methodological change but also a source of significant risk. Since then the situation has grown more complex: AI is no longer just a tool, it is increasingly shaping and organizing the research process itself.
AI now enters the realm of scientific decision-making: it can propose research questions, search for data sources, define variables, suggest models, write and debug code, interpret results, and edit text. Researchers therefore receive not just an answer, but a potential research pathway — and that changes what we regard as relevant, convincing, or scientific.
Causal reasoning versus prediction
A core lesson of economics is that correlation is not causation. AI systems are strong at pattern recognition and prediction, but such outputs do not automatically provide explanation: they may fail to grasp institutions, historical context, or societal consequences. For that reason, causal thinking becomes even more important as a defense against superficially convincing but ultimately weak scientific claims.
This institutional sensitivity is particularly necessary for Hungarian research topics: regional inequalities, the social integration of Roma communities, performance of education and health institutions, changes in state capacity, or regional differences in firm productivity cannot be fully understood from general model-derived patterns alone.
The cost of productivity: new validation burdens
AI can increase researcher productivity in some tasks — generating code drafts, organizing structured tables, preparing literature overviews, or improving language. Yet it also introduces new verification costs: when AI speeds up context-dependent tasks, researchers may gain time in production but spend more time validating and justifying decisions.
This is especially consequential for doctoral training: if early-career researchers no longer learn from faulty code, poorly specified models, contradictory literature, or failed empirical strategies, the craft of research may be weakened. The correct response is not blanket bans or unrestricted use, but training that teaches doctoral students to distinguish execution from understanding — they may use AI for coding or document preparation, but must justify model choices, variable construction, identification strategies and interpretation of results.
Infrastructure and language: unequal access
AI depends on substantial infrastructure: data centers, chips, energy, cloud services, privacy systems, licensing agreements and institutional capacity. Not all researchers or institutions have equal access to these resources. In Hungary this also involves a language dimension: large language models are mostly trained on English-language material, which forces Hungarian researchers into double mediation — they must interpret reality and also interpret how the machine translates that reality into a global scientific language. Linguistic sovereignty here means ensuring that Hungarian institutional and historical experience can still be processed in its own terms.
Domestic inequalities matter too: large Budapest universities and research centers have different opportunities than smaller regional institutions — access to data, computing power and legal support vary. AI does not automatically democratize science; without adequate institutional conditions it may widen gaps between center and periphery, rich and poor institutions, and well-supported and unsupported researchers.
Data as an institutional product
Data are not raw commodities but institutional products: someone collected, defined, cleaned, documented and regulated them. Hungarian empirical economics illustrates this: analyses based on KSH microdata, controlled-access research facilities, or datasets curated by the KRTK Adatbank require knowledge of survey logic, variable meanings, breaks in time series, legal access conditions and privacy constraints. AI can aid documentation processing and accelerate coding, but it cannot substitute for the expertise that explains what a variable signifies in the Hungarian institutional context.
Research attention and selection effects
AI tends to favor data-rich, well-structured, easily formalizable topics. Individually this incentivizes researchers to choose such problems — faster results, more papers — but at the system level it can distort the research agenda. Slower, harder-to-quantify but socially crucial issues may be sidelined. In Hungary many important problems fall into that category: persistent poverty, low trust and rule-following, dysfunctional public administration, the social costs of economic transitions, regional lagging or the role of informal institutions.
AI is therefore a selection mechanism, not a neutral magnifier: it brings some problems closer and pushes others further away. This bias must be recognized.
Tensions in the publication system
AI's entry into scholarly publishing has mixed effects. It can improve language quality and help with technical preparation, which benefits non-native English speakers. Yet it may also increase manuscript submissions while peer-review capacity remains limited. Journals and scholarly communities thus need to distinguish forms of AI use (language editing, code writing, data processing, literature search, hypothesis generation, result interpretation) and require documentation of decision points.
Responsibility and traceability
One of AI's greatest risks is not merely that it makes mistakes, but that the origin of mistakes becomes blurred. Decision points — who defined a variable, who chose a model specification, why alternative explanations were omitted — must be traceable and justifiable. AI is not an author in the professional sense: it does not bear professional responsibility, defend claims in debate, assess societal consequences, or be held accountable for misguided policy advice. Responsibility remains with the researcher. Thus transparency is necessary but not sufficient: traceability of decisions, documentation of checks and substantive human sign-off are equally essential.
Conclusion: transformation rather than replacement
AI does not render economists obsolete, but it reshapes their role: researchers become organizers, verifiers and interpreters of the research process. That is not a lesser function; it demands greater responsibility — theoretical rigor, causal reasoning, institutional sensitivity and social responsibility.
For the Hungarian economics community the task is to avoid either uncritical adoption of international methodological fashions or reflexive rejection of new tools. Instead, build on national research traditions, data infrastructure and Hungarian-language knowledge so that AI serves rather than subsumes the logic of science. Machines work fast, but they cannot decide in which direction that speed should take science — that decision remains human.
Fertő Imre is Director General of the Institute of Economics (KTI) at ELTE Közgazdaság- és Regionális Tudományi Kutatóközpont (KRTK) and a professor at Budapesti Corvinus Egyetem (BCE). The article reflects the author’s opinion and does not necessarily represent the editorial stance of Portfolio.
