Advances in Artificial Intelligence: Implications for Capital Markets Activities

After completing this reading, you should be able to:

• Describe current uses of artificial intelligence (AI) and machine learning (ML) in capital markets, and potential future uses of sophisticated AI models including GenAI.
• Explain the implications of further adoption of AI on market dynamics.
• Explain the implications of further adoption of AI on financial stability.
• Describe best practices for regulators pertaining to AI, including the use of AI by supervisors and recommendations for policy on the use of AI by regulated entities.

Uses of Artificial Intelligence (AI) and Machine Learning (ML) in Capital Markets

Artificial intelligence and machine learning are already integrated into the investment lifecycle of capital markets, though their visibility is less than popular stories about fully automated finance suggest. In reality, most current uses enhance existing analytical and operational processes instead of replacing human judgment. Financial institutions have relied on statistical learning and algorithmic tools for many years, with recent progress mainly focused on scale, data integration, and computing power. These tools now enable firms to handle larger and more complex datasets, improve the consistency of analysis, and decrease manual efforts in areas where traditional models face challenges with high dimensionality or unstructured data.

One of the most pervasive current uses of AI is transforming diverse data sources into actionable investment signals. Machine learning models are applied to both structured financial data and unstructured sources such as earnings calls, regulatory filings, news, and alternative datasets. Natural language processing (NLP) techniques convert text into structured indicators by measuring sentiment, identifying themes, and detecting changes in tone and emphasis over time. These indicators are then incorporated into existing investment frameworks rather than used in isolation, improving the informational content of traditional signals without discarding established valuation or factor-based models.

Example: NLP-Driven Signal Extraction in Equity Research

Consider an equity long-short fund that supplements its fundamental research with an ML model trained on earnings call transcripts and corporate disclosures. The model detects shifts in management language related to capital expenditure plans and pricing power, flagging companies whose narrative tone diverges from historical patterns or sector peers. Analysts use these flags to prioritize further investigation rather than to adjust positions automatically. In this way, AI enhances research coverage, timeliness, and screening efficiency while preserving human oversight and interpretive judgment.

These extracted signals feed into security selection and relative value analysis, where AI models identify nonlinear relationships across assets and markets. Techniques such as clustering and network analysis help reveal evolving correlations, crowding effects, and hidden factor exposures that may not be apparent in traditional linear models. This is particularly valuable in complex portfolios, where diversification benefits can unexpectedly erode as market conditions change.

Example: Detecting Hidden Correlation Risk

A multi-asset portfolio manager relies on an ML-based clustering model to monitor correlation structures across equity, credit, and commodity exposures. During a period of rising inflation uncertainty, the model indicates that several assets previously considered diversifying are increasingly driven by the same macro factor. Although headline correlations remain modest, the clustering analysis reveals growing concentration risk. The manager responds by adjusting exposures before the correlations fully materialize during a market stress episode.

At the portfolio level, AI and ML support forecasting, risk estimation, and portfolio optimization in high-dimensional settings. Advanced models help estimate time-varying expected returns, volatilities, and covariances, improving the robustness of portfolio construction under uncertainty. Rather than redefining allocation frameworks, these tools refine inputs and stress assumptions, particularly in multi-asset strategies where interactions among risk factors are complex and unstable.

AI also plays a central role in trade execution and market interaction. Execution algorithms use ML to assess liquidity conditions, forecast short-term price movements, and dynamically adjust execution strategies to minimize transaction costs and market impact. In less liquid markets, AI supports price discovery by combining sparse transaction data with related instruments and market signals to infer executable prices.

Example: AI-Optimized Trade Execution

An institutional investor executing a large corporate bond trade uses an AI-driven execution system that analyzes recent trades, dealer quotes, and correlated instruments to estimate fair value and optimal execution timing. The system recommends splitting the order and executing through multiple venues to reduce price impact. Traders approve the strategy and monitor execution, intervening only if market conditions shift unexpectedly.

Beyond front-office activities, AI is increasingly used to enhance liquidity management and operational efficiency. Machine learning models forecast margin calls and collateral needs by analyzing portfolio volatility, counterparty behavior, and historical stress patterns. This allows firms to manage funding proactively rather than reactively, particularly during periods of market turbulence.

Example: Liquidity Forecasting and Margin Management

A clearing member applies an ML model to forecast potential variation margin calls under different volatility scenarios. When the model signals a high likelihood of increased margin requirements, the treasury team adjusts liquidity buffers and funding plans in advance. This reduces the risk of forced asset sales during market stress.

Monitoring, compliance, and reporting represent another central area of current AI use. Machine learning systems detect anomalous trading patterns, flag potential market abuse, and automate compliance screening. AI-driven tools also support risk reporting by generating tailored dashboards and summaries for internal risk committees and regulators, improving transparency and consistency while still requiring human validation.

Finally, AI is increasingly used by supervisors and in regulatory technology (RegTech) and supervisory technology (SupTech) applications. Supervisory authorities employ AI-based tools for market monitoring and early risk detection, while banks use similar techniques to enhance anti-money laundering (AML), know-your-customer (KYC), and fraud detection processes. These applications influence market behavior indirectly by shaping surveillance intensity, compliance expectations, and operational resilience across institutions.

Taken together, current uses of AI and ML form a tightly interconnected ecosystem across data analysis, investment decision support, execution, liquidity management, and oversight. The defining feature of current adoption is augmentation rather than autonomy: AI improves speed, scale, and analytical depth, but humans retain responsibility for interpretation, control, and accountability. This foundation sets the stage for more sophisticated models to alter not just how information is processed, but how decisions themselves are generated.

As AI models become more advanced, their potential uses extend beyond support toward decision generation and adaptive portfolio management. Advances in model architecture, computing power, and generative artificial intelligence (GenAI) enable systems not only to identify patterns but to produce investment narratives, strategy proposals, and scenario interpretations. This represents a qualitative shift with implications for governance, accountability, and systemic risk.

One important potential development lies in investment research and strategy formation. GenAI models can synthesize large volumes of structured and unstructured information to generate investment theses and scenario-based insights, shifting the analyst’s role from information gathering toward critical evaluation.

Example: GenAI-Assisted Investment Thesis Generation

A global asset manager deploys a GenAI system trained on historical macroeconomic cycles and policy responses. When inflation surprises persist, the system generates alternative scenarios linking monetary policy paths, real interest rates, and asset-class performance, proposing portfolio tilts consistent with each narrative. Portfolio managers review and challenge the outputs before deciding which scenarios merit implementation.

Sophisticated AI models may also enable end-to-end portfolio management, linking forecasting, allocation, execution, and risk monitoring into unified decision engines capable of near-real-time adjustments. While this integration improves responsiveness, it increases reliance on model behavior under stress and raises the risk that errors propagate rapidly across the investment process.

Example: End-to-End AI-Driven Portfolio Adjustment

In a multi-asset strategy, an advanced AI system monitors macro indicators, market prices, and liquidity conditions simultaneously. Following an unexpected policy announcement, the system reallocates exposures, adjusts execution timing, and updates liquidity buffers without direct human intervention. Risk managers oversee the system through escalation thresholds, but response speed exceeds manual capabilities.

A particularly significant future application lies in stress testing and scenario analysis. GenAI models may generate forward-looking stress scenarios that extend beyond historical precedents, capturing nonlinear transmission channels and systemic feedback effects across markets and institutions. While these scenarios enhance tail-risk awareness, their opacity complicates validation and interpretation.

Example: AI-Generated Stress Scenarios

A systemically important institution uses a GenAI model to generate stress scenarios combining geopolitical shocks, commodity disruptions, and funding market stress. Risk teams use these scenarios to test capital and liquidity resilience, even though precise probabilities cannot be assigned.

Generative AI may also reshape market communication, disclosure, and regulatory interaction, while advanced supervisory uses could enhance systemic risk monitoring and policy simulation. However, widespread reliance on similar models raises concerns around herding, synchronized behavior, and shock amplification.

Overall, the evolution from current AI applications to sophisticated, generative systems marks a shift from augmentation to partial autonomy. As AI increasingly influences how strategies are formed and executed, the core risk management challenge shifts toward model governance, human oversight, and clear accountability, ensuring that technological advances enhance market efficiency without undermining financial stability.

Implications of Further Adoption of AI on Market Dynamics

The further adoption of artificial intelligence in capital markets has implications that extend beyond individual firm efficiency to the behavior and structure of markets as a whole. As AI systems increasingly influence how information is processed, trades are executed, and risks are managed, they also shape price formation, liquidity conditions, and volatility dynamics. These effects arise not only from what AI systems do individually, but from how many market participants adopt similar technologies simultaneously, often trained on overlapping data and optimized toward comparable objectives.

One immediate implication of broader AI adoption is its effect on market reactions to news and price discovery. AI systems can process information faster and incorporate new data into prices more rapidly than human-driven processes. This can improve informational efficiency in normal conditions by reducing mispricing and narrowing arbitrage opportunities. However, faster information incorporation also means that markets may adjust more abruptly to new signals, compressing reaction times and leaving less scope for discretionary intervention. As a result, price movements may become sharper and more discontinuous, particularly when multiple AI-driven strategies respond simultaneously to the same information.

Example: Accelerated Price Adjustment

Following a surprise macroeconomic data release, multiple asset managers using AI-based signal-extraction models revise their return forecasts within seconds. Equity index futures and bond yields adjust sharply as algorithms rebalance portfolios almost simultaneously. While prices quickly reflect the new information, intraday volatility spikes because the adjustment occurs in a compressed time window rather than being spread out over hours.

AI adoption also affects liquidity provision and withdrawal. Under stable conditions, AI-driven execution algorithms can enhance liquidity by optimizing order placement, narrowing bid–ask spreads, and improving matching efficiency. However, during periods of stress, the same systems may rapidly reduce exposure or withdraw from markets if models detect rising risk or deteriorating liquidity. Because many algorithms rely on similar signals, this can lead to sudden liquidity evaporation, amplifying price movements and market fragility.

Example: Algorithmic Liquidity Withdrawal

During a volatility surge triggered by geopolitical news, AI-based market-making algorithms detect abnormal order-book behavior and rising adverse selection risk. The systems automatically widen spreads or suspend quoting altogether. Although each algorithm acts rationally from a risk management perspective, the collective withdrawal leads to a sharp decline in market depth and exaggerated price swings.

Another important implication concerns volatility amplification and feedback loops. AI systems often rely on signals derived from recent price movements, volatility measures, or market flows. When markets move sharply, these inputs can trigger reinforcing responses, such as deleveraging, stop-loss execution, or risk-based rebalancing. This creates endogenous volatility, where market movements are driven not only by external news but also by the internal dynamics of AI-driven strategies reacting to one another.

Example: Volatility Feedback Mechanism

A sudden equity sell-off increases realized volatility measures used by several AI-driven risk parity and volatility-targeting strategies. In response, these systems reduce equity exposure, adding further selling pressure. The resulting price decline feeds back into volatility estimates, triggering additional deleveraging and amplifying the original shock.

Widespread AI adoption also raises concerns about herding behavior and model convergence. Although AI systems may appear diverse, many are trained on similar datasets, use comparable architectures, and are optimized against similar performance metrics. This can lead to correlated decision-making across institutions, even in the absence of explicit coordination. Herding driven by AI differs from traditional behavioral herding in that it can occur rapidly, mechanically, and at scale, making it harder to detect and counteract in real time.

Beyond trading behavior, AI influences market structure and competitive dynamics. Firms with access to superior data, computing infrastructure, and technical expertise may gain persistent advantages, potentially increasing concentration among large institutions and technology providers. Smaller participants may become more dependent on third-party AI services, improving operational interdependencies and creating common points of failure. These structural shifts can alter who provides liquidity, how risks are distributed, and how shocks propagate through the system.

Example: Concentration and Dependency Risk

Several mid-sized asset managers outsource AI-driven analytics to the same third-party provider. When the provider experiences a system outage or model error, multiple firms simultaneously lose access to critical signals, impairing their ability to trade or manage risk. The disruption affects market activity more broadly because of the shared dependency.

Finally, further AI adoption complicates market transparency and interpretability. As decision-making becomes more model-driven, it becomes harder for market participants, regulators, and even firms themselves to explain why certain trades occurred or why liquidity disappeared at specific moments. Reduced transparency can undermine confidence during stress events and make it more difficult for authorities to assess whether markets are functioning in an orderly manner or whether intervention is required.

In summary, increased AI adoption reshapes market dynamics by accelerating information transmission, altering liquidity behavior, amplifying volatility through feedback mechanisms, and increasing correlation across participants. While these effects can enhance efficiency in normal times, they also raise the risk that markets become more fragile under stress. For risk managers and regulators, the key challenge lies in understanding not just individual AI models, but the collective behavior that emerges when many such models interact within tightly connected financial markets.

Implications of Further Adoption of AI on Financial Stability

The expanding adoption of artificial intelligence in capital markets has implications that go beyond market dynamics to the stability of the financial system as a whole. Financial stability concerns arise when the collective behavior of institutions threatens the orderly functioning of markets, the resilience of key intermediaries, or the continuity of critical financial services. As AI systems increasingly influence decision making across trading, risk management, and liquidity provision, they introduce new channels through which shocks can be amplified, propagated, or synchronized across institutions and markets.

A central financial stability concern relates to the speed and scale of shock transmission. AI systems operate at speeds far beyond human reaction times, allowing positions to be adjusted, liquidity to be withdrawn, and exposures to be rebalanced almost instantaneously. While this speed enhances efficiency under normal conditions, it also compresses adjustment periods during stress. When adverse information arrives, AI-driven responses may occur simultaneously across multiple institutions, leaving little room for stabilizing measures or discretionary intervention. As a result, localized shocks can escalate rapidly into system-wide stress events.

Example: Rapid Shock Propagation

Following an unexpected sovereign downgrade, AI-based risk models across banks and asset managers simultaneously revise credit risk assessments. Automated systems reduce exposures to affected assets within minutes, triggering sharp price declines and spillovers into related markets. Although each institution acts prudently from a micro risk perspective, the collective speed of response amplifies the shock and strains market functioning.

Another key channel is liquidity risk amplification. AI systems are increasingly playing a role in liquidity provision through algorithmic trading, execution optimization, and collateral management. In stable markets, these systems improve liquidity efficiency. However, during stress, AI-driven liquidity providers may withdraw abruptly if models detect elevated volatility, adverse selection, or funding pressure. Because many institutions rely on similar liquidity signals, withdrawal can occur in a correlated manner, increasing the likelihood of liquidity freezes and fire-sale dynamics.

This dynamic is particularly concerning for financial stability because liquidity shortages often interact with leverage and margin requirements. When asset prices fall and volatility rises, AI models forecasting margin calls and funding needs may trigger defensive actions, including asset sales and balance-sheet contraction. These responses can reinforce price declines, creating self-reinforcing liquidity spirals.

Example: AI-Driven Liquidity Spiral

During a period of market stress, rising volatility causes AI models at clearing members to forecast higher variation margin requirements. Firms preemptively liquidate assets to raise cash, increasing selling pressure and further depressing prices. The resulting volatility feeds back into margin models, triggering additional liquidity demands and asset sales across the system.

The further adoption of AI also raises concerns about model convergence and systemic herding. Although AI systems may appear diverse, many are trained on overlapping datasets, optimized using similar loss functions, and constrained by comparable regulatory and risk-management frameworks. This can lead to correlated behavior across institutions, even in the absence of explicit coordination. From a financial stability perspective, such herding reduces the diversity of responses to shocks, weakening the system’s natural shock-absorbing capacity.

AI-driven herding is particularly dangerous because it can be fast, mechanical, and opaque. Unlike traditional behavioral herding, which may unfold gradually, AI-driven alignment can occur almost instantaneously, making it difficult for supervisors to detect emerging systemic risks before they materialize.

Example: Convergent Risk Reduction

Multiple asset managers employ volatility-sensitive AI models to manage portfolio risk. When volatility rises modestly, all models reduce exposure to the same asset class within a short time window. The resulting synchronized selling overwhelms market depth, transforming a manageable fluctuation into a systemic stress event.

Another important financial stability implication relates to procyclicality. AI systems often rely on historical data to estimate risk, volatility, and correlations. During benign periods, these models may encourage increased leverage and risk-taking by signaling low risk and stable conditions. Conversely, during downturns, rapidly rising risk estimates can force widespread deleveraging. This reinforces the financial cycle, making expansions more aggressive and contractions more severe.

Procyclicality is not new, but AI can intensify it by increasing the responsiveness and uniformity of risk signals across institutions. As AI models react quickly to changing inputs, they may accelerate balance-sheet adjustments that destabilize markets rather than smooth them.

The use of AI also introduces new operational and infrastructure risks with systemic implications. Structural dependencies created by AI adoption pose stability risks. Increased interconnectedness and concentration risks arise as financial institutions rely on common cloud infrastructure, shared data providers, and third-party AI platforms. This reliance creates common points of failure, where operational disruptions, outages, or governance weaknesses at a small number of providers can simultaneously affect many institutions. At the same time, the high fixed costs associated with advanced AI systems favor large, well-resourced firms. As market activity becomes more concentrated among a small group of dominant participants, the distress or failure of any one of them can have outsized systemic consequences.

Example: Third-Party Dependency Shock

Several major market participants rely on the same cloud-based AI platform for trading analytics and risk monitoring. A system outage during a volatile trading session prevents firms from accessing real-time signals and executing risk controls. The disruption impairs liquidity provision and increases uncertainty across markets, exacerbating instability.

Opacity and model interpretability risk further complicate financial stability oversight. As AI systems become more complex, it becomes increasingly difficult for institutions and regulators to understand how decisions are made or how models will behave under extreme conditions. This limits the ability to anticipate systemic interactions, validate stress responses, and intervene effectively during crises. Reduced transparency can undermine confidence precisely when trust in market functioning is most critical.

Deeper reliance on AI may weaken traditional stabilizing mechanisms within financial institutions. Erosion of human oversight and governance risks occurs as automation reduces the role of discretionary judgment in trading, risk management, and operational decision making. Human intervention has historically acted as an informal circuit breaker by questioning model outputs, delaying action, or reassessing assumptions during periods of stress. As decision pathways become shorter and more automated, these buffers may weaken. Without strong governance frameworks, clear accountability, and effective override mechanisms, technical failures or model-driven errors can escalate rapidly into broader financial disruptions. From a financial stability perspective, further adoption of AI primarily acts as a powerful risk amplifier, intensifying existing vulnerabilities rather than introducing entirely new forms of risk.

Finally, the interaction between AI adoption and regulatory frameworks may itself influence financial stability. If regulatory constraints, stress tests, or capital rules embed similar AI-driven metrics across institutions, they may inadvertently reinforce correlated behavior. Conversely, insufficient oversight of AI use may allow risks to accumulate unnoticed until they crystallize abruptly.

It can therefore be concluded that further adoption of AI reshapes financial stability risk by increasing speed, interconnectedness, and correlation across the financial system. While AI can enhance risk management at the individual firm level, it may simultaneously weaken stability at the system level if collective effects are not well understood and managed. The core challenge for regulators and risk managers is therefore not whether AI improves decision-making locally, but whether its system-wide interactions increase or reduce the resilience of the financial system under stress.

Best Practices for Regulators on AI Use in Capital Markets

Given the rapidly evolving and uncertain landscape of AI in capital markets, regulators are encouraged to prioritize engagement and outreach as a first line of action rather than rushing into rigid enforcement. The IMF notes that supervisors have so far leaned more toward clarification and outreach than enforcement, even as many jurisdictions develop broader AI strategies and consider dedicated AI legislation. This stance reflects a practical reality: supervisors must understand how AI is being deployed, where risks concentrate, and whether existing frameworks remain fit for purpose before prescribing highly specific rules.

A core best practice is the creation of structured mechanisms for dialogue with industry. The IMF highlights establishing public–private forums to develop overarching principles, partnering with industry to build a risk framework, and conducting surveys to assess how existing frameworks apply to AI. These tools serve two purposes. First, they help authorities map where AI is actually used across services and activities. Second, they allow supervisors to test whether existing risk management guidance sufficiently captures AI-specific challenges, such as explainability, robustness, data bias, privacy, and cybersecurity.

As part of oversight expectations, regulators are encouraged to treat AI governance as an extension of technology-neutral supervisory logic, while still recognizing that AI can introduce distinctive model risk features. The IMF notes that existing regulatory and supervisory frameworks in capital markets are largely technology-neutral and applicable to AI systems, with ongoing work exploring whether additional frameworks are needed to address risks specific to AI use, including conduct issues such as ethics, fairness, and transparency. This is consistent with broader principles emphasized by standard setters, including results-based and proportional regulation and supervision, which aim to capture risks without freezing innovation.

A second pillar of best practice is for supervisors to strengthen their own capacity and toolkits by actively adopting SupTech. The IMF notes that AI can generate efficiency gains for supervisors by automating data quality checks for completeness, correctness, and consistency, and by combining multiple datasets even when unique identifiers are missing. It also highlights supervisory use cases such as anomaly detection in trading patterns reflected in changes in prices, volume, and volatility, as well as efforts to identify misleading information and support real-time monitoring of market transactions. These capabilities directly address a practical supervisory constraint: AI-driven markets can produce volumes and speed of activity that exceed what traditional monitoring can reliably process.

Example: SupTech for Market Surveillance

A supervisor deploys an AI model that continuously scans market data for unusual clustering of price moves and volume spikes across venues. When the system detects patterns consistent with coordinated algorithmic behavior, supervisors can prioritize targeted investigation and request additional information from relevant firms and brokers. This aligns with the IMF’s emphasis on anomaly detection, real-time monitoring, and improved data integration for oversight.

GenAI further expands supervisory possibilities by enhancing information retrieval, content creation, and code generation and debugging, including modernization of legacy code. In supervisory settings, this matters because it can speed up the deployment of traditional oversight functions such as fraud detection and monitoring of market activity while also streamlining data management tasks. However, adopting these tools safely requires supervisors to invest in periodic upskilling, both to understand AI risks and to detect behaviors such as models designed to game the system or forms of algorithmic coordination.

A complementary best practice is to use cross-sectoral oversight approaches to detect systemwide effects that are not visible at the level of a single firm. The IMF points to the value of cross-sectoral thematic reviews to identify potential herding and material interconnectedness among participants, and to surface best practices in AI use across the market. This is especially relevant when AI adoption creates correlated strategies, shared data dependencies, or common third-party model providers, all of which can become systemic vulnerabilities.

From a policy perspective, the IMF emphasizes that regulation and supervision in AI-related areas should be enhanced to address potential financial stability risks across banks and nonbank financial intermediaries, using a balanced approach that balances benefits with risks. Market participants themselves expect regulators to provide clarity and guidance on model risk management, emphasize stress testing for extreme scenarios, and strengthen transparency and disclosures, while avoiding overly rigid rulemaking given the speed of technological change.

A critical supervisory recommendation for regulated entities is to require regular risk mapping of interdependencies between data, models, and technological infrastructure supporting AI systems. This mapping is essential because AI models may rely on shared architectures and a small number of software, data, and cloud providers, and because datasets may not span a complete financial cycle, undermining reliability under stress. The IMF also highlights that while frameworks often require assessing cumulative effects of models, they may not mandate a joint assessment of data dependencies, making explicit mapping a practical supervisory upgrade.

Regulators are also encouraged to enhance monitoring and data collection for market participants whose activities have an outsized impact. This includes strengthening oversight of nonbank financial intermediaries by requiring them to identify themselves and disclose AI-relevant information, and monitoring participants that conduct substantial trading activity, often referred to as large traders, using unique identification and reporting mechanisms. These measures improve visibility into fast-moving, technology-driven market activity and reduce the likelihood of risk build-ups in opaque corners of the system.

Example: Large Trader Monitoring

A capital markets authority introduces a reporting regime requiring firms that exceed a trading activity threshold to be uniquely identified and provide information on their algorithmic strategies through their broker-dealers. This improves supervisory ability to link abnormal price and volume patterns to specific sources of activity and strengthens real-time risk monitoring.

Where AI adoption increases market speed and volatility under stress, the IMF points to policy actions that focus on market guardrails rather than model design alone. Authorities and trading venues should assess whether new or modified volatility response mechanisms are necessary, including whether circuit breakers require recalibration as market structures evolve. It also notes that testing algorithms in controlled environments can help authorities and market actors assess behavior in extreme circumstances. Alongside this, trading venues and central counterparties should review margining requirements and buffers given the potential for rapid AI-driven price moves.

Another major policy area involves operational resilience and third-party dependencies. The IMF recommends a coordinated approach to the regulation and supervision of critical AI service providers by mapping relationships between critical AI providers and essential IT infrastructure providers, and by ensuring that the definition of critical providers is broad enough to capture the systemic use of common models. Comparable and interoperable approaches facilitate compliance and coordination across authorities, especially when major providers serve multiple institutions simultaneously. In parallel, authorities should strengthen cyber resilience by requiring protocols to prevent, detect, respond to, and recover from attacks, recognizing that AI systems can be attacked through both training data manipulation and model extraction.

Finally, regulators should pay attention to areas where opacity is structurally higher and where AI could accelerate fragility, such as over-the-counter (OTC) markets. The IMF recommends preparedness measures to maintain market integrity and resilience, including collecting and disseminating more detailed OTC transaction information, requiring participants to account for liquidity shifts in their risk frameworks, improving incentives for market making and central clearing, and establishing margin requirements for non-centrally cleared derivatives. These policy tools are designed to preserve resilience even if AI adoption increases speed, interconnectedness, and the likelihood of correlated liquidity withdrawal.

Note that best practices for regulators combine proactive engagement, capability building, and targeted policy responses. Supervisors should stay vigilant and modernize their tools to monitor more complex strategies and process granular data in real time, while ensuring regulated entities strengthen governance through transparency, risk mapping, and robust model risk management. The practical objective is not to prevent AI adoption, but to ensure that AI’s efficiency gains do not translate into greater systemic fragility through opacity, concentration, procyclicality, and operational vulnerabilities.

Practice Question

A regulator considers allowing financial institutions to use GenAI systems to automatically generate stress scenarios that combine geopolitical shocks, commodity supply disruptions, and funding market stress, even when no historical precedent exists. Probabilities are not assigned, but scenarios are used to test resilience. What feature most clearly distinguishes this potential future use of AI from current stress testing practices?

1. Reliance on backward-looking historical simulation
2. Reduction in the need for capital and liquidity buffers
3. Elimination of model risk through broader datasets
4. Ability to construct forward-looking, non-historical scenarios

Correct Answer: D

This example highlights a key potential future use of GenAI: generating forward-looking stress scenarios that are not constrained by historical data. Traditional stress testing typically relies on past crises, stylized shocks, or expert-designed scenarios. GenAI systems can instead synthesize information across domains and construct plausible but novel stress narratives that capture nonlinear transmission channels and systemic feedback effects. While these scenarios improve tail-risk awareness, they also raise challenges related to validation, interpretability, and governance, since probabilities cannot be reliably assigned and model behavior under extreme conditions is harder to assess.

A is incorrect. The defining feature here is moving beyond backward-looking simulation.
B is incorrect. Stress testing does not reduce the need for buffers; it informs their adequacy.
C is incorrect. Broader data does not eliminate model risk and may increase opacity.

Things to Remember

• GenAI can extend stress testing beyond historical experience.
• Forward-looking scenarios improve tail-risk exploration.
• Validation and interpretability become more challenging.