Central banks are turning to artificial intelligence to analyse data and improve economic forecasts. But Elisa Mustacchi argues that this raises a question that goes beyond accuracy and efficiency. If AI increasingly shapes the information behind monetary-policy decisions, who remains accountable for the judgments that follow?
Artificial intelligence is already part of the analytical machinery of central banking. At the European Central Bank, AI helps collect, clean and analyse large amounts of data. The ECB applies machine-learning tools to inflation forecasting and to assessing inflation risks in real time.
These uses may sound technical, but they have an institutional importance because monetary policy depends on information. The final decision may remain entirely human, but as AI becomes more deeply embedded in the analytical chain, the question of accountability begins well before that decision is taken.
Central banks have always relied on models to understand the economy. What AI changes is the speed, scale and variety of information those models can process.
But not all uses of AI are equally consequential. There is a difference between using an algorithm to classify millions of data points and using the algorithm to produce forecasts, simulate scenarios or generate recommendations on which economists and policymakers may rely. The formal authority to decide can remain entirely human while AI plays an increasingly important role in shaping the evidence presented to decision-makers.
There is a difference between using an algorithm to classify millions of data points and using the algorithm to produce forecasts, simulate scenarios or generate recommendations on which economists and policymakers may rely
This is the reason why the distinction between decision support and decision-making is less reassuring than it first appears. The more a system influences which trends, scenarios or risks a bank considers relevant, the more it can shape the reasoning that follows.
AI does not need to make a decision to influence one.
The obvious response is to keep a human in the loop. But this phrase can conceal more than it clarifies. Human oversight is meaningful only if the people involved can understand, question and, when necessary, reject what the system produces.
In practice, this means more than approving a final recommendation. Experts need to validate models, check the consistency of their outputs, identify possible distortions in the data and interpret results in their broader economic context. An AI system may detect a correlation or flag a risk, but we still need human judgment to determine whether that signal is plausible, relevant or misleading.
This matters because we cannot hold an algorithm accountable for a monetary-policy judgment. It cannot defend a decision publicly or assume responsibility for an error. The ECB itself has stressed that humans must remain firmly in control, but in control of what? Linking human control to accountability and public trust, without looking at the interpretation given by the AI, is not necessarily meaningful human control.
This question is particularly important for central banks because independence gives them considerable discretion over decisions with wide economic consequences. But independence does not mean freedom from scrutiny. The ECB describes accountability as the necessary counterpart to its independence: it must explain its actions and policies to citizens and their elected representatives, primarily through the European Parliament.
If an AI's forecasts or risk assessments become influential inputs into monetary-policy analysis, the reasoning behind a decision may increasingly depend on systems that are difficult to reconstruct or explain
AI complicates this relationship without formally changing it. A model does not need legal authority to affect institutional reasoning. If its forecasts or risk assessments become influential inputs into monetary-policy analysis, part of the reasoning behind a decision may increasingly depend on systems that are difficult to reconstruct or explain.
The problem, then, is not simply the prospect of an algorithm making an autonomous decision. Responsibility can also become harder to locate as algorithmic outputs move through the analytical process. Excessive reliance on opaque models could make it more difficult to explain how particular assessments were reached, even if responsibility for the final decision formally remains clear.
Independence therefore makes traceable responsibility more important, not less.
If AI increasingly shapes the evidence on which monetary-policy decisions are based, accountability cannot begin only when the final decision is taken. It must extend back through the analytical process.
This does not mean that every model or technical output needs to be made public. It means that when AI plays a material role in an analysis, the institution should be able to reconstruct how that contribution entered the decision-making process: what the system produced, who reviewed the output, what limitations were identified, whether the result was questioned or overridden, and who ultimately assumed responsibility for using it.
If AI increasingly shapes the evidence on which monetary-policy decisions are based, accountability cannot begin only when the final decision is taken
Without this traceability, responsibility can become diluted across models, technical teams and policymakers. Formal responsibility may remain perfectly clear while substantive responsibility becomes much harder to identify.
Accountability should therefore follow the influence of AI, and not simply apply to formal decision-making authority.
AI can help central banks process more information, improve forecasts and identify risks more quickly. These are significant gains. But efficiency alone does not make a public decision legitimate.
As AI becomes more deeply embedded in central-bank analysis, the challenge is to ensure that its contribution remains open to human scrutiny and that responsibility never disappears behind the model.
The challenge, then, is not simply to keep humans in the loop. It is to keep responsibility in the loop.
A fascinating and well written insight. You mention "linking human control to accountability and public trust" is important. But who has ever been held accountable for faulty models? The ECB is still accountable as an institution and will remain so. It is an institution using multiple tools and multiple persons' analyses. None of these are accountable per se. It is expected that forecasts and predictions will be wrong to some extent. Furthermore the assumptions embedded in the models are political (despite many economists protestations). Thus, AI or not, much of the forecasting and scenario testing of monetary policy is made to support policy decisions rather than shape them.
It is useful you take up the issue of accountability using the paradigm of AI input. Accountability and knowledge is surely a central part of what a democratic society is based on. Is there transparency in models used by central banks? Does ECB offer current models such as the NAWM in a form that lets outsiders reconstruct them, tweek them or completely modify them?
Perhaps turning your argument on its head, what do you think if AI was used to validate human produced models? AI could, misquoting you, "reconstruct what contribution entered the decision-making process: what the system produced, who reviewed the output, what limitations were identified, whether the result was questioned or overridden, and who ultimately assumed responsibility for using it".
Thank you — this is a very thoughtful comment. I agree that the ECB ultimately remains institutionally accountable, whether it uses AI, traditional models, or human analysis. My concern is therefore less about making a model or an individual analyst “accountable” for a forecast that turns out to be wrong, and more about preserving traceability within an increasingly complex decision-making process.
Forecasts will inevitably be wrong to some extent, and, as you point out, model assumptions are themselves never entirely neutral. The question for me is whether we can still reconstruct how a particular piece of analysis influenced a policy decision, what assumptions and limitations were known, and how competing evidence was assessed.
Your point about transparency is particularly important. The ability of outsiders to reconstruct, challenge or modify central-bank models is closely connected to meaningful accountability, regardless of whether AI is involved.
I also really like your suggestion of turning the argument around. AI could potentially be valuable precisely as a tool for challenging human-built models — identifying assumptions, inconsistencies or sensitivities that might otherwise be overlooked. The interesting question then becomes how we validate the validator: if AI is used to scrutinise human reasoning, its own assumptions, limitations and outputs would also need to remain contestable and transparent.
Thank you again — definitely food for thought.