How is AI changing economics and forecasting?
We began by discussing a simple taxonomy of artificial intelligence (AI) which presented the overlapping nature of machine learning, deep learning, natural language processing and their intersection in the form of large language models (LLMs), which currently dominate the AI conversation. With this in mind, we looked at AI’s contribution in three broad areas: improving economic measurement, enhancing forecasting and reshaping the balance between prediction and explanation.
During the roundtable, participants agreed that AI in all its forms is expanding measurement possibilities, offering powerful tools for reducing the cost of processing large volumes of structured and unstructured data – including text and other non-traditional data sources. This enables economists to construct “nowcasts” of economic activity, sentiment and business conditions, while also improving research productivity through the automation of data pipelines.
With regard to forecasting, machine learning techniques have shown they can outperform traditional methods, particularly in environments characterised by complex, non-linear relationships and large numbers of predictors. However, because these methods are often viewed as “black boxes”, participants questioned how much economic judgement should continue to inform forecasts and how machine learning models should complement, rather than replace, traditional time-series approaches. With respect to LLMs, participants saw them primarily as tools to support forecasters by reducing processing time and freeing up capacity to challenge assumptions, test new ideas and explore alternative scenarios. The broad consensus was that LLMs are unlikely to replace forecasters but instead will augment their work.
Perhaps the most fundamental debate centred on the relationship between prediction and explanation. Given that economics has traditionally focused on understanding causal mechanisms, we raised the question whether increasingly powerful AI models might lead to economists trading interpretability for predictive accuracy, or whether explicability must remain central to economic analysis and policy advice. The general consensus was the latter.
How might AI be incorporated into macroeconomic modelling?
Our second topic was whether recent advances in AI might fundamentally reshape macroeconomic modelling. We began by asking whether AI could help overcome one of the long-standing challenges facing heterogeneous-agent (HA) models: their computational complexity. In particular, could AI make high-dimensional models more tractable, and could it allow economists to explore a much richer, more realistic range of behavioural assumptions and market frictions than is currently feasible?
This question resulted in considerably less agreement. Some argued that LLM-powered agents could offer a more realistic representation of decision-making than the representative-agent frameworks that continue to dominate much of macroeconomics. Others were more sceptical, noting that LLMs do not remove the fundamental dependence of agent-based models on behavioural assumptions. Simply increasing the number or sophistication of agents does not necessarily produce better economics if the underlying behavioural rules remain poorly founded.
One of the strongest themes to emerge was the perceived promise of reinforcement learning in which agents learn through interaction, feedback and repeated experience rather than relying on fixed behavioural rules. This was seen as providing a useful middle ground between the unrealistic assumption of perfectly rational expectations and entirely ad hoc behavioural assumptions.
Several participants argued that learning and imperfect information are fundamental features of real economies that traditional macroeconomic models often struggle to capture. Standard models frequently assume that agents possess complete knowledge of the economic environment, leaving little need for communication or learning over time. AI-based approaches, particularly reinforcement learning, may offer a more realistic way of modelling how expectations evolve through experience.
At the same time, it was repeatedly emphasised that data remains a fundamental constraint. Reinforcement learning and many other AI techniques require extensive data for training, while LLMs implicitly rely on enormous datasets that have already been absorbed during pre-training. However, they cautioned against viewing LLMs as “assumption-free”: they embed assumptions just as conventional models do, but those assumptions are often less transparent.
The discussion ultimately returned to a familiar question in macroeconomics: what constitutes success? Most agreed that AI is unlikely to replace existing macroeconomic frameworks outright, but it offers promising new tools for modelling learning, heterogeneity and complex interactions. The central challenge will be finding an appropriate balance between realism, interpretability, computational feasibility and empirical validation.
What methodological questions are raised by the application of AI model?
The final topic broadened the focus from macroeconomic models to the economics profession itself; exploring how AI may change the way economists conduct research, communicate evidence and contribute to policymaking.
A recurring concern throughout the discussion was transparency. Economists and policymakers have traditionally placed considerable value on models that are interpretable and whose underlying mechanisms can be understood and scrutinised. By contrast, many AI systems operate as black boxes, making it difficult to explain exactly how particular outputs are generated. Participants questioned how much confidence policymakers should place in recommendations that cannot be fully explained or independently verified.
Related concerns arose around reproducibility. Unlike traditional statistical analysis, current LLM-generated outputs vary depending on the prompts given, the LLM used (e.g. ChatGPT or Claude) or which version of the same LLM was employed. Participants recognised that this creates new challenges for transparency, replication and scientific credibility.
The discussion also considered how the role of economists might evolve as AI increasingly automates many research tasks. While coding, literature reviews and some forms of data analysis may become substantially easier, economists will still have a role to play. They will increasingly be expected to interrogate and defend AI-generated outputs, challenge questionable conclusions, exercise professional judgement and communicate evidence effectively to policymakers. The consensus was that the future is likely to involve AI complementing the work of economists rather than substituting for it.
Participants also highlighted potential implications for the profession itself. Several expressed concern that larger institutions with deeper pockets may gain an increasing advantage because they possess the resources, data and computing infrastructure to develop and deploy sophisticated AI systems. Smaller universities and research institutions could struggle to keep pace, potentially widening inequalities in research capacity across the profession.
Finally, the discussion turned to the development of future generations of economists. In the early stages of their career, they traditionally acquired expertise through hands-on experience in collecting data, building models and writing code. If AI increasingly performs these tasks, there is a risk that future generations may become highly proficient at using AI tools without developing the deeper understanding needed to critically evaluate their outputs. It was stressed that maintaining opportunities to develop economic intuition and technical expertise will remain essential.
Despite these concerns, the overall tone remained cautiously optimistic. There was general agreement that AI has the potential to transform economic research by automating routine tasks and expanding analytical capabilities. At the same time, interpretation, economic reasoning, communication and sound judgement remain fundamentally human activities. The discussion concluded that, in an AI-enabled world, the comparative advantage of economists will increasingly lie not in producing analysis, but in understanding, questioning and applying it.
This blog is based on discussion held at the ‘AI in Macroeconomics Roundtable‘ event.