BIS: The Use of Artificial Intelligence for Policy Purposes

Why the BIS Is Mapping AI Adoption Across Central Banks

The Bank for International Settlements, often described as the central bank of central banks, published a landmark report in October 2025 titled “The Use of Artificial Intelligence for Policy Purposes.” The document represents the most comprehensive mapping of AI experimentation across the global central banking community to date, drawing on insights from dozens of institutions and multiple BIS Innovation Hub projects.

Central banks face a unique challenge. They must process enormous volumes of economic data, monitor complex financial systems in real time, and make policy decisions that affect billions of people — all while maintaining strict standards for accuracy, transparency, and accountability. The BIS report argues that artificial intelligence, and in particular machine learning and large language models, can fundamentally enhance how central banks fulfill these mandates.

The report organizes AI applications into four interconnected policy domains: information collection and statistics, macroeconomic analysis and monetary policy, payments oversight and anti-money laundering, and supervision and financial stability. Each domain features practical pilots, measurable results, and candid assessments of what works and what does not. For anyone tracking how AI is reshaping financial regulation, this BIS publication is essential reading.

What makes this report distinct from typical technology surveys is its grounding in operational reality. Every use case discussed has been tested, either through BIS Innovation Hub proofs-of-concept like Aurora, CB-LM, and Gaia, or through direct experimentation by central banks in both advanced and emerging economies. The report does not merely theorize — it provides implementation blueprints.

How AI Transforms Data Collection and Statistical Analysis

Central banks are custodians of vast microdata repositories spanning banking transactions, credit registries, securities holdings, and economic surveys. Traditionally, cleaning, validating, and compiling this data has been a labor-intensive manual process. The BIS report documents how machine learning algorithms — particularly isolation forests and ensemble methods — now automate outlier detection, error flagging, and data quality assurance at a scale impossible for human analysts alone.

Isolation forests, a tree-based anomaly detection technique, have proven especially effective for identifying suspicious or erroneous entries in large financial datasets. Rather than requiring predefined rules for what constitutes an anomaly, these models learn the structure of normal data and flag observations that deviate significantly. Central banks using these tools report dramatic reductions in manual review time while simultaneously catching more subtle data quality issues.

Beyond cleaning, AI enables entirely new data sources. Web scraping combined with natural language processing allows central banks to track online prices in real time, creating high-frequency inflation indicators that complement traditional CPI surveys. Satellite imagery analysis can estimate economic activity — crop yields, port traffic, nighttime illumination — offering nowcasting capabilities that were science fiction a decade ago. The BIS Innovation Hub has supported several of these data innovation projects across its centers in Singapore, Hong Kong, and Basel.

The implications extend beyond efficiency. Better data quality translates directly into better policy decisions. When a central bank can identify and correct reporting errors in real time rather than months later, its inflation estimates, credit risk assessments, and financial stability indicators all become more reliable. This foundational improvement in data infrastructure is arguably the most immediately deployable AI benefit documented in the report.

Machine Learning for Macroeconomic Forecasting and Monetary Policy

Monetary policy has always depended on forecasts — of inflation, GDP growth, employment, and financial conditions. Traditional econometric models use structured, low-frequency data published with significant lags. The BIS report shows how machine learning is transforming this forecasting landscape by incorporating unstructured, high-frequency data sources and non-linear modeling approaches.

Neural networks and gradient-boosted tree models can process text from news articles, social media posts, earnings call transcripts, and central bank speeches to generate real-time sentiment indices. These sentiment series have demonstrated predictive power for consumer spending, business investment, and financial market stress — variables that traditional models capture only with delay. The European Central Bank and the Federal Reserve have both published research on incorporating text-derived variables into nowcasting frameworks.

Few-shot and zero-shot learning capabilities of modern large language models add remarkable flexibility. A central bank economist can now prompt a pre-trained model to classify economic news by theme, extract inflation expectations from survey responses, or summarize hundreds of regional economic reports — tasks that previously required dedicated teams and custom models. The BIS notes that these capabilities are particularly valuable for emerging market central banks with limited data science resources.

However, the report cautions against over-reliance. Machine learning models can capture complex patterns but may also amplify noise, particularly in volatile periods when historical patterns break down. The recommended approach combines ML-generated signals with traditional structural models, using ensemble methods where the strengths of each approach compensate for the weaknesses of others.

AI-Powered Anti-Money Laundering: The Aurora Project

Perhaps the most striking case study in the entire BIS report is Project Aurora, a proof-of-concept developed by the BIS Innovation Hub’s Nordic Centre. Aurora tested whether machine learning — specifically graph neural networks — could detect complex money-laundering schemes that traditional rule-based systems consistently miss.

Current AML compliance relies heavily on predefined transaction thresholds and pattern rules. Criminals have learned to structure transactions just below these thresholds, creating layered networks of shell accounts and nominees that are nearly invisible to rule-based systems. The result is a paradox: banks spend billions on AML compliance while detecting only a small fraction of actual laundering activity, and the overwhelming majority of flagged transactions turn out to be false positives.

Aurora took a fundamentally different approach. Instead of analyzing individual transactions against fixed rules, graph neural networks map the entire network of payment flows, identifying suspicious patterns in the relationships between accounts. A single transaction may appear benign, but when viewed within the context of a broader network — circular flows, unusual counterparty clustering, timing patterns — the same transaction becomes a strong laundering indicator.

The results were dramatic. Aurora’s network-centric ML approach significantly outperformed traditional methods in detecting complex laundering typologies, including those involving multiple jurisdictions and numerous intermediary accounts. Critically, false positive rates dropped substantially, meaning investigators could focus their limited resources on genuinely suspicious activity rather than processing thousands of spurious alerts.

The privacy dimension deserves emphasis. Aurora explored privacy-preserving techniques — federated learning and secure multi-party computation — that allow multiple banks to train a shared model without exposing their raw transaction data to each other or to a central authority. This addresses one of the fundamental barriers to cross-institutional AML: the tension between effective detection (which requires seeing the full network) and data protection (which restricts data sharing).

Supervision and Financial Stability Through Natural Language Processing

Financial supervisors review enormous volumes of documents: annual reports, regulatory filings, board minutes, risk disclosures, audit opinions, and correspondence from supervised entities. The BIS report details how NLP tools now automate and augment much of this document processing, enabling supervisors to identify emerging risks faster and allocate examination resources more strategically.

Sentiment analysis applied to banks’ earnings calls and public communications can provide early warning signals of financial stress. When the tone of a bank’s management shifts — becoming more hedging, more defensive, or more focused on liquidity — NLP models detect these shifts before they manifest in hard financial data. Several supervisory authorities have built dashboards that aggregate these text-derived risk signals across their entire supervised population.

Topic modeling and document classification further enhance efficiency. Instead of manually routing thousands of regulatory filings to the correct review teams, ML classifiers can automatically categorize documents by risk theme, flag those requiring urgent attention, and even extract specific data points from unstructured text. The BIS reports that some central banks have reduced document processing times by over 70% using these techniques.

For financial stability analysis, AI supports more sophisticated stress-testing scenarios. Traditional stress tests use a handful of predetermined macroeconomic scenarios. Machine learning enables scenario generation at scale, creating thousands of plausible adverse scenarios that capture tail risks and interaction effects that human-designed scenarios might miss. Combined with network analysis of interbank exposures, these tools provide a richer picture of systemic vulnerabilities.

Large Language Models Fine-Tuned for Central Bank Operations

One of the most technically sophisticated findings in the BIS report concerns domain-specific fine-tuning of large language models. The CB-LM (Central Bank Language Model) project demonstrated that taking a general-purpose foundation model and fine-tuning it on central banking texts — speeches, working papers, policy statements, and research publications — produces materially better results on domain-specific tasks than using the base model alone.

The improvements span multiple applications. For sentiment analysis of monetary policy communications, CB-LM achieved significantly higher accuracy in classifying hawkish versus dovish statements compared to both general-purpose LLMs and earlier NLP approaches. For document summarization, the fine-tuned model generated more accurate and contextually appropriate summaries of economic research. For question-answering over regulatory texts, CB-LM demonstrated stronger comprehension of specialized terminology and institutional context.

The principle extends beyond central banking. Any domain with a substantial corpus of specialized text — legal, medical, regulatory — can benefit from fine-tuning foundation models rather than relying on general-purpose capabilities. The BIS positions this finding as one of the most actionable recommendations: central banks should invest in building and maintaining domain-specific training corpora and fine-tuned models, potentially sharing them across institutions to reduce duplication of effort.

The BIS Irving Fisher Committee has already begun coordinating efforts to create shared datasets and benchmarks for evaluating AI models on central banking tasks. These collaborative infrastructure investments could significantly lower the barrier to AI adoption for smaller central banks that lack the resources to develop models independently.

Privacy-Preserving Technologies for Cross-Border Collaboration

One of the most significant barriers to effective AI deployment in the financial sector is data fragmentation. Money laundering crosses borders, systemic risk spans institutions, and inflation dynamics connect economies — but the data needed to analyze these phenomena is locked in organizational and jurisdictional silos, protected by legitimate privacy and sovereignty requirements.

The BIS report devotes substantial attention to privacy-enhancing technologies (PETs) that promise to unlock collaborative AI without compromising data protection. Three approaches receive particular emphasis: federated learning, secure multi-party computation, and encrypted aggregation.

Federated learning allows multiple institutions to train a shared machine learning model without ever exchanging raw data. Each institution trains the model locally on its own data, then shares only the model updates (gradients) with a central coordinator that aggregates them. The resulting model benefits from the combined data of all participants while no single institution ever sees another’s data. The Aurora project used this approach for cross-bank AML analytics.

Secure multi-party computation (SMPC) takes a different mathematical approach, enabling multiple parties to jointly compute a function over their private inputs without revealing those inputs to each other. For financial applications, SMPC could allow banks to compute aggregate exposure statistics or correlations without any bank revealing its individual positions.

The report is candid about remaining challenges. PETs add computational overhead, they require standardized data formats and agreed protocols, and their legal status in many jurisdictions is still ambiguous. The Financial Stability Board and BIS are working on establishing common frameworks, but implementation at scale is likely years away. Nonetheless, the report positions PETs as essential infrastructure for the next generation of financial sector AI applications.

Managing AI Risks: Explainability, Hallucinations, and Vendor Concentration

The BIS report does not shy away from the substantial risks that AI adoption introduces. Five interconnected risk categories receive detailed treatment, and together they form the basis for the governance framework recommendations in the final section.

Explainability and transparency. The most accurate machine learning models — deep neural networks, large ensembles — are often the least interpretable. For central banks, which must justify their decisions to legislatures, markets, and the public, this opacity creates a fundamental tension. A model that perfectly predicts a banking crisis is of limited policy value if no one can explain why it predicted the crisis or what factors drove the prediction.

Hallucinations. Large language models sometimes generate confident, detailed, but entirely fabricated outputs. In a consumer chatbot, a hallucination is embarrassing; in a monetary policy briefing or supervisory assessment, it could be catastrophic. The BIS emphasizes that human-in-the-loop validation remains non-negotiable for any high-stakes AI application, regardless of how impressive the model’s average accuracy may be.

Data governance and privacy. Legal uncertainty around web scraping, intellectual property rights for training data, and cross-border data transfer regulations creates compliance risks for central banks that deploy AI tools trained on public data. Several jurisdictions are actively litigating these questions, and the regulatory landscape could shift dramatically.

Vendor concentration. The AI supply chain is remarkably concentrated. A handful of companies provide the cloud infrastructure, GPU capacity, and foundation models that most AI applications depend on. If a major provider experiences an outage, security breach, or business discontinuity, the operational impact could cascade across the financial system. The BIS recommends that financial authorities treat large AI providers as part of the critical infrastructure perimeter, subject to due diligence, diversification requirements, and contingency planning.

Cyber threats. AI is a dual-use technology in cybersecurity. While it strengthens defensive capabilities — better anomaly detection, faster incident response — it also empowers attackers. Generative AI can produce more convincing phishing emails, deepfake voice calls for social engineering, and sophisticated prompt-injection attacks against AI-integrated systems. The cybersecurity implications demand that central banks significantly increase their security investments alongside their AI investments.

Building an AI Governance Framework for the Financial Sector

The final and arguably most impactful section of the BIS report synthesizes the technical findings into a comprehensive governance framework for AI in financial policy. The framework addresses strategy, operations, talent, regulation, and international cooperation.

Strategy and institutional governance. Central banks should establish explicit AI strategies that include model risk management frameworks, ethics guidelines, procurement policies, and audit trails. Every AI system deployed in production should have documentation covering its training data, methodology, known limitations, validation results, and incident response procedures. The report recommends adopting FAIR principles (Findable, Accessible, Interoperable, Reusable) for all data and model metadata.

Operational resilience. AI systems should be treated with the same rigor as other critical infrastructure. This means independent validation teams, regular auditing, red-team testing, and clearly defined “kill switches” that can disable AI components without disrupting core operations. Vendor risk management must evolve to address the unique dependencies introduced by AI providers.

Talent and organizational change. The report identifies workforce transformation as one of the most challenging aspects of AI adoption. Central banks need cross-disciplinary teams that combine domain expertise (economics, law, supervision) with technical skills (data science, ML engineering, cybersecurity). Using the unique appeal of public-service missions and access to extraordinary data, central banks can compete for talent against the private sector — but only if they modernize their hiring practices, career structures, and technical infrastructure.

Phased deployment. The BIS recommends a two-phase approach. In the near term, deploy “AI copilots” that augment human analysts — drafting documents, summarizing data, flagging anomalies — while keeping humans firmly in control of all decisions. In the medium term, as governance matures and trust accumulates, consider more autonomous AI agents for narrowly defined, well-monitored tasks.

International cooperation. Perhaps most importantly, the report calls for scaled-up international collaboration through the BIS, the Financial Stability Board, and the G20. AI governance cannot be effective at the national level alone because the technology, the data, the providers, and the risks are all global. Cross-jurisdictional standards for model validation, provider oversight, and data sharing are essential — and the BIS positions itself as the natural coordinator for this effort.

For organizations looking to make their own research and policy documents more accessible and engaging, platforms like Libertify offer a compelling solution. Rather than distributing static PDFs that often go unread, institutions can transform their publications into interactive experiences that increase comprehension, engagement, and ultimately impact.