Where Should Economists Draw the Line in Using AI? 

By Imre FertÅ‘ 

AI can make research faster, but the scarce resource in economics will increasingly be judgement, validation and responsibility. 

Artificial intelligence (AI) is no longer merely a new tool on the economist's desk. Less than a year ago, I argued in an earlier AES blog post that AI was transforming economic research by expanding the range of data, methods and analytical possibilities available to researchers. The central message was that economists should embrace these opportunities, but only under the disciplines of rigour, transparency, causal reasoning and institutional responsibility. 

That argument still stands. But the debate has moved on. 

The most important question today is not whether economists should use AI. They already do. Nor is the central question whether AI will replace economists. It will not replace the distinctive role of economic reasoning, at least not if the profession remains clear about what that role is. The deeper question is where responsibility lies when AI becomes involved not only in executing research tasks, but also in shaping the research process itself. 

AI can now help formulate research questions, identify data sources, define variables, suggest empirical strategies, write code, interpret results and polish manuscripts. This is a profound change. The economist no longer receives only an answer to a narrowly defined question. Increasingly, the economist is offered a possible research pathway. 

That makes AI more than a productivity tool. It becomes part of the intellectual infrastructure of research. 

From result validation to workflow validation 

Economists are used to checking results. Is the dataset appropriate? Is the specification defensible? Are the estimates robust? Is the causal interpretation credible? These questions remain essential. But AI adds another layer of responsibility: economists must also examine how the research process arrived at those results. 

Who suggested the research question? Why was one dataset chosen over another? Which variables were constructed, omitted or redefined? Which alternative methods were not considered? What assumptions were built into AI-generated code or text? 

In other words, causal validation must be complemented by workflow validation. The economist must be able to reconstruct and defend not only the final estimate, but also the chain of decisions that produced it. 

This is especially important in agricultural and applied economics. Consider a study of agri-environmental schemes, farm productivity, climate adaptation or rural inequality. AI can help process satellite imagery, administrative records, farm accountancy data or large bodies of policy documents. It can also summarise literature, suggest hypotheses and draft empirical code. But it cannot, by itself, understand why a particular subsidy works differently across regions, why farmers respond differently to the same incentive, or how institutional trust shapes policy take-up. 

Those are not merely technical issues. They are economic and institutional questions. For example, the same agri-environmental payment may produce different outcomes in two regions not because farmers calculate incentives differently in a narrow economic sense, but because advisory services, administrative capacity, previous experiences with public programmes and trust in local institutions differ. AI may detect the pattern, but the economist must explain the institutional mechanism behind it. 

Prediction is not explanation 

AI is extraordinarily powerful at detecting patterns. This is one of its great strengths. But it is also the source of a familiar danger for economists: prediction can be mistaken for explanation. 

A model may predict yields, land-use changes or food price movements with impressive accuracy. Yet policymakers often need to know more than what is likely to happen. They need to know why it happens, under what institutional conditions it happens, and what would change if a policy instrument were altered. 

This distinction is central to economics. Correlation is not causation; predictive performance is not policy understanding. A machine-learning model may detect that certain farms are more likely to adopt a climate-smart practice. But the policy question is whether a change in incentives, information, regulation or trust would cause adoption to increase. That requires theory, institutional knowledge and credible identification. 

AI can assist that work. It can make the research process faster and broader. But it cannot relieve economists of the responsibility to ask causal questions. 

The hidden cost of productivity 

Much discussion of AI begins with productivity. In some tasks, the gains are real. AI can draft code, clean text, organise references, translate documents, summarise reports and improve the readability of academic writing. For researchers working in a second language in particular, this can reduce barriers to international publication. 

But productivity is not free. AI also creates new verification costs. 

When the task is well structured and the output is easy to check, AI may save time. When the task requires deep contextual judgement, the gain is less obvious. A complex empirical strategy, a sensitive interpretation of institutional change or a policy conclusion based on heterogeneous effects cannot be delegated without loss. The researcher may save time in production, but lose time in validation. 

The bottleneck therefore shifts. In many cases, the scarce resource will no longer be the capacity to generate text, code or preliminary analysis. It will be the capacity to evaluate, interpret and take responsibility for them. 

This has implications for doctoral training. Young researchers learn not only from correct answers, but also from failed models, messy data, rejected hypotheses and contradictory literatures. If AI removes too many of these learning moments too early, it may weaken the craft of research. The right response is not to ban AI, nor to allow its uncritical use. It is to teach doctoral students to distinguish execution from understanding. 

A student may use AI to help write code or structure a literature review. But they must still be able to justify the model, the variables, the identification strategy and the interpretation of the findings. 

Data are not raw material 

Economists often speak of data as if they were a natural resource waiting to be extracted. This is misleading. Data are institutional products. They are defined, collected, cleaned, documented, regulated and made accessible by people and organisations. 

This matters greatly for empirical research. Farm-level data, household surveys, administrative records and environmental indicators all carry histories of measurement and classification. A variable does not become economically meaningful simply because an algorithm can process it. It becomes meaningful when the researcher understands what it represents in a specific institutional and social context. 

AI can help navigate documentation and detect inconsistencies. But it cannot replace knowledge of how data are produced, what they omit, and how their institutional origins shape interpretation. 

This is particularly important for research outside the dominant English-language academic core. Local-language policy documents, legal categories, administrative practices and historical experiences may not be fully represented in global AI systems. Translation can help, but it can also flatten concepts. Economists must therefore protect not only methodological rigour, but also institutional and linguistic sensitivity. 

AI and the direction of scientific attention 

AI may also affect what economists choose to study. If AI provides the greatest advantage in data-rich, highly structured and easily formalised topics, researchers may have stronger incentives to focus on those areas. At the individual level, this is understandable. Such topics may be faster to analyse and easier to publish. 

At the system level, however, the consequence may be a narrowing of scientific attention. Problems that are slower, less structured or less data-rich may receive less attention, even when they are socially important. 

In agricultural economics, some of the most important questions are precisely of this kind: the resilience of small farms, informal land-use practices, intergenerational change in rural communities, trust in public institutions, or the local implementation of environmental policies. These questions may not always fit neatly into machine-readable formats. But that does not make them less important. 

AI is not a neutral magnifying glass. It can bring some problems into sharper focus while pushing others to the margins. Economists should be aware of this selection effect. 

Responsibility cannot be delegated 

The use of AI in research should be transparent. Yet transparency alone is not enough. It is not sufficient to state that AI was used. We need to know where it was used, for what purpose, and under whose judgement. 

There is a difference between using AI for language editing, coding assistance, data cleaning, literature search, hypothesis generation and interpretation of results. The closer AI comes to the core scientific claim, the greater the responsibility of the researcher. 

The central danger is not simply that AI may make mistakes. Human researchers also make mistakes. The danger is that the origin of the mistake may become unclear. Who chose the specification? Who ignored an alternative mechanism? Who overstated the policy conclusion? 

If the economist cannot answer these questions, they have lost control of their own research. 

AI is not an author in the economic sense. It cannot defend a claim, accept professional responsibility, consider social consequences or be held accountable for flawed policy advice. Responsibility remains with the economist. 

The economist's role is changing, not disappearing 

AI will not make economists unnecessary. But it will change what responsible economic research requires. The economist will become less of a mere executor of analysis and more of a designer, supervisor and interpreter of the research process. 

That is not a lower-status role. It requires higher-level judgement. 

Economists must understand the logic of AI without becoming subordinate to it. They must preserve theoretical discipline, causal reasoning, institutional knowledge and social responsibility. The speed of the machine cannot determine the direction of science. 

AI can make economic research faster. It can make some forms of analysis richer and more accessible. But it can also distort incentives, increase unchecked claims, widen institutional inequalities and blur the boundaries of responsibility. 

The task, therefore, is not to reject AI. It is to ensure that the logic of the tool does not replace the logic of the discipline. 

Economics is not only about arranging data and running models. It is about understanding economic and social systems, identifying causal mechanisms and assessing consequences responsibly. AI can assist in that work. But it cannot replace the judgement that makes economics a science. 

This post develops arguments from my recent Hungarian article on economists' responsibility in the use of AI and follows my earlier AES blog post on artificial intelligence and economic research.