CI-2 The Financial Stability Implications of Artificial Intelligence

AI DEVELOPMENTS SINCE 2017: OVERVIEW

The FSB first assessed AI in financial services in 2017. The 2024 update identifies four headline developments that were nascent or nonexistent at that time:

DEEP LEARNING

Multi-layered neural networks now handle complex, unstructured data with high accuracy, enabling pattern recognition at scale across text, images, and voice.

BIG DATA

The volume and variety of available data has grown dramatically. Firms now process satellite images, social media feeds, and transaction records simultaneously.

COMPUTE POWER

GPUs now handle mass parallel calculations cheaply, making previously cost-prohibitive model training feasible for large financial institutions.

GENERATIVE AI & LLMS

ChatGPT (Nov 2022) marked a step-change. LLMs allow ordinary language interaction with machines, dramatically lowering barriers to AI adoption across financial services.

Exam tip: The FSB distinguishes supply-side (primary driver since 2017) vs demand-side drivers. Supply-side factors have driven the most significant changes.

SUPPLY-SIDE DRIVERS OF AI ADOPTION

Supply-side factors represent what became technologically and commercially possible. The FSB groups these into three categories:

TECHNOLOGICAL DEVELOPMENTS

• Deep learning & embeddings: Better handling of unstructured data
• Transformer architecture: Foundation of LLMs, attention mechanism enables context understanding
• GPU hardware: Parallel computation at scale, AI investment projected $400B by 2027

DATA-RELATED DEVELOPMENTS

• Digitization of customer interactions generated vast unstructured data (text, images, video)
• Synthetic data emerging as workaround to real-data exhaustion (potentially by 2026)
• Growing GenAI model complexity demands ever-larger training datasets

BUSINESS MODEL DEVELOPMENTS

• Cloud computing enables flexible access to storage, compute, and AI toolsets
• Pre-trained models allow deployment without training from scratch
• In 2023, 66% of new foundation models were open-source (vs 33% in 2022)

DEMAND-SIDE DRIVERS OF AI ADOPTION

The FSB identifies three demand-side drivers that have remained consistent since 2017. They have not materially changed but continue to reinforce AI uptake.

PROFITABILITY

AI improves profitability through four channels:

• Revenue generation (better targeting, reduced churn)
• Cost reduction (automating back-office tasks)
• Risk management (predicting delinquencies)
• Operational efficiency

COMPETITION

• Fear of falling behind peers is arguably the strongest demand-side pressure.
• FIs that fail to adopt AI risk losing customers, market share, and talent to competitors.

REGULATORY COMPLIANCE

Increasing regulatory requirements across jurisdictions:

• AML/CFT screening & monitoring
• Know-Your-Customer (KYC) processes
• Data protection rules

AI USE CASES IN FINANCIAL SERVICES: INDUSTRY OVERVIEW

The FSB uses an activity-based framework. Be able to classify a given use case by type and recognize which AI techniques apply.

Key caveat: Actual customer-facing AI adoption remains low (under 6.7% of US financial firms as of mid-2024). Survey data is skewed toward large firms.

INDUSTRY USE CASES: FOUR CORE CATEGORIES

CUSTOMER-FOCUSED

• Credit underwriting: ML supplements traditional credit scores
• Insurance pricing: ML replaces generalized linear models
• Chatbots & virtual assistants for customer service

OPERATIONS-FOCUSED

• Capital optimisation: ML for regulatory capital modeling
• Volatility & liquidity management under stress scenarios
• Code generation & document automation

TRADING & PORTFOLIO MANAGEMENT

• LLMs analyze sentiment from earnings calls & filings
• Reinforcement learning for trade execution optimization
• Automated investment research & portfolio rebalancing

REGULATORY COMPLIANCE (REGTECH)

• Wider AML/CFT use cases: sanctions evasion, trade-based money laundering
• GenAI automates financial crime investigation reports
• Tax evasion detection across jurisdictions

REGULATORY & SUPERVISORY USE CASES (SUPTECH)

Financial authorities are themselves adopting AI to supervise FIs more effectively. 59% of surveyed authorities used AI-based supervisory tools in 2023, up 5pp from 2022.

CURRENT SUPTECH APPLICATIONS

• NLP & GenAI for analyzing earnings calls, filings, & management discussions
• Extracting key paragraphs from inspection documents
• Drafting supervisory summaries

KEY CONSTRAINT: SKILLS GAP

64 financial regulators reported organisational skills deficiencies in data science & IT.

This matters because:

• Supervisory effectiveness suffers
• FI adoption may outrun oversight capacity

The asymmetry: FIs are building AI capabilities faster than regulators can evaluate them, creating a monitoring gap that is itself a financial stability risk.

Exam angle: SupTech is both a solution and itself subject to data, governance, and skills constraints.

THIRD-PARTY DEPENDENCIES & CONCENTRATION

VULNERABILITY 1: MOST STRUCTURALLY COMPLEX

Arises from three interdependent dynamics: increasing use of AI, concentration of AI supply chains, and criticality of the services those chains enable.

1  HARDWARE DEPENDENCY

AI models require GPUs & ASICs for training and deployment. Design and fabrication markets dominated by very few firms.

FIs either purchase chips directly or rent capacity through cloud providers.

2  CLOUD CONCENTRATION

CSPs supply compute, host LLMs, and provide API access. Major CSPs both train & distribute models.

FIs incur indirect cloud exposure through specialized AI service providers.

3  MODEL DEPENDENCY

Training LLMs from scratch is cost-prohibitive. Most FIs rely on pre-trained models from few providers.

Vendor-provided risk models for fraud, credit scoring, & cyber defence.

Vertical integration amplifies risk: Certain firms provide hardware, software, cloud, and models simultaneously, creating concentration across all layers.

LLM MARKET CONCENTRATION: FORCES IN BOTH DIRECTIONS

The FSB concludes: AI infrastructure exhibits increasing returns to scale. The critical policy question: could a single provider failure become a systemic event?

THE LLM SUPPLY CHAIN

KEY INPUTS TO MODEL PRODUCTION

• GPUs: Most concentrated segment. Few firms dominate design & fabrication
• Data: Open-source web crawls, proprietary, synthetic
• Compute: Cloud service providers or direct GPU access

THREE DISTRIBUTION CHANNELS

• Direct to end-user: Curated products (chatbots, code tools) or developer APIs
• Open repositories: Pre-trained models for download/customization (98+ text gen models by early 2024)
• Cloud-hosted: CSPs offer LLMs as managed services, creating cloud dependency

Greatest concentration risk: Hardware (GPU design & fabrication). Model training is more competitive. Cloud access is intermediate risk.

MARKET CORRELATIONS

VULNERABILITY 2

When many FIs use the same AI models and training data, predictions and strategies converge. This invisible herding creates correlated risk.

WHY MODELS BECOME CORRELATED

• Herding: firms imitate peers’ model choices
• Network externalities: models improve with more users
• Limited choice: few models meet performance bar
• Data convergence: similar training data used

FINANCIAL STABILITY IMPACT

• Amplified volatility in trading
• Exacerbated liquidity crunches
• Flash crash probability increased
• Shared credit model failures cause synchronized lending contractions

AUTOMATION AS AN AMPLIFIER

AI correlations are most dangerous in highly automated markets. Semi-automated treasury systems reacting in real time could create synchronized deposit outflows.

MITIGATING PATH

If AI enables truly customized investment strategies, it could reduce correlations. Net effect depends on differentiation vs common-platform efficiency.

CYBER RISKS

VULNERABILITY 3

AI changes the cyber threat landscape in two ways: it empowers attackers & increases the attack surface.

AI EXPANDS ATTACKER CAPABILITIES

• Social engineering: convincing phishing, voice impersonation, deepfakes
• Business email compromise: automated fraudulent comms
• Code generation: GenAI writes exploit code, lowers skill barrier
• Reconnaissance: automated vulnerability scanning

AI EXPANDS THE ATTACK SURFACE

• Data poisoning: manipulate training data for wrong outputs
• Model poisoning: alter weights directly
• Adversarial inputs: exploit model vulnerabilities
• Prompt injection: hijack LLM behavior

Short-run asymmetry: Malicious actors adopt GenAI without guardrails. In the near term, attackers may benefit more from AI breakthroughs than defenders.

MODEL RISK, DATA QUALITY, & GOVERNANCE

VULNERABILITY 4

AI introduces features that make model risk harder to manage than in traditional quantitative models. Three issues combine.

LIMITED EXPLAINABILITY

Deep learning & LLM models are opaque. When a model cannot explain its reasoning:

• Independent validation is constrained
• Supervisors cannot evaluate
• Bias detection becomes harder
• Accountability for errors is unclear

DATA QUALITY & OPACITY

Modern AI trains on massive, heterogeneous datasets FIs are unaccustomed to evaluating:

• LLM training data is opaque/undisclosed
• Standard data quality checks cannot apply
• Non-traditional data creates governance uncertainty

GOVERNANCE GAPS

Low-cost, easy-deploy AI leads to rapid adoption without controls:

• FIs skip validation, testing, oversight steps
• Existing model risk frameworks may not cover GenAI
• Hallucinations in customer-facing LLMs create liability

FRAUD, DISINFORMATION, & MISALIGNMENT

VULNERABILITY 5

Three additional risks tied specifically to GenAI capabilities & AI behavior in complex environments.

FRAUD

• Synthetic identity creation: fake IDs, voice clones, deepfakes bypass KYC
• False insurance claims fabricated with AI-generated evidence
• Scale of losses rising globally

DISINFORMATION

• Fake images of bank closures trigger social-media bank runs
• Deepfake videos of CEOs making market-moving announcements
• Fabricated financial reports or data

MISALIGNMENT

AI systems optimize their objective without regard for unintended consequences:

• Trading AI may engage in market manipulation inadvertently
• Lending AI may systematically exclude protected groups
• Autonomous agents may take actions beyond intended scope

LONGER-TERM STRUCTURAL CONSIDERATIONS

Three longer-term dynamics affecting financial stability. Not near-term vulnerabilities, but exam-relevant for systemic context.

COMPETITIVE LANDSCAPE

If custom AI is key, large FIs gain structural advantage. If off-the-shelf GenAI wins, smaller FIs can compete. Outcome determines consolidation pressure.

MACROECONOMIC CONDITIONS

GenAI threatens high-skill cognitive work (unlike prior automation of routine tasks). Labour market disruptions could affect output, wages, inflation, and financial stability.

ENERGY USE

AI accounts for ~1% of global energy consumption, growing. Model inference may exceed training energy as end-user applications scale.

FINANCIAL STABILITY VULNERABILITY SUMMARY

WHAT THE FSB RECOMMENDS

Existing frameworks address many but not all AI vulnerabilities. AI developments require additional policy considerations.

ADDRESS DATA GAPS

Surveys, reporting, disclosure
regimes for FI & AI provider
engagement to fill data gaps
on AI adoption patterns

ENHANCE REGULATORY
CAPABILITIES

International cooperation among
financial authorities, share best
practices, leverage SupTech tools
for supervision

ASSESS FRAMEWORK
ADEQUACY

Evaluate domestic & international
regulatory frameworks for fitness
against AI-specific risks that
existing rules may not cover

Key exam point: No single jurisdiction or authority can address these issues alone. The cross-border nature of AI supply chains requires international coordination among financial authorities, data regulators, and privacy bodies.