Artificial Intelligence (AI)

Artificial intelligence (AI) is the branch of computer science concerned with building systems that perform tasks normally associated with human intelligence: recognizing patterns, understanding language, making decisions and learning from experience. The term covers a wide range of methods, from fixed rules written by experts to models with billions of parameters trained on large bodies of data.

In financial services, AI is used to detect fraud, score credit, read documents, answer customers, and support trading and investment decisions. Most systems in use are narrow: each does one job and has no general understanding. Banks, asset managers and regulators treat AI models as a source of model risk that needs testing, documentation and human oversight like any other model.

The idea that a machine could think was posed in practical form by Alan Turing in 1950, in a paper that proposed a test based on whether a person could tell a machine's written answers from a human's. The name "artificial intelligence" was introduced by John McCarthy for a research workshop held at Dartmouth College in 1956.

Early work centered on logic and hand-written rules. The 1980s brought expert systems, which stored the knowledge of specialists as if-then rules and were used by banks for tasks such as loan assessment. Progress came in waves separated by periods of reduced funding known as AI winters. From about 2012, deep learning produced large gains in image and speech recognition, helped by faster chips and bigger data sets. The public release of ChatGPT in November 2022 made generative AI a mass-market product.

• Rule-based systems. People write the logic directly. The system is easy to explain but cannot handle cases its authors did not foresee.
• Machine learning. In machine learning the system finds its own rules from examples. It may learn from labeled data, from unlabeled data, or by trial and reward, the last being reinforcement learning.
• Neural networks. An artificial neural network stacks many simple units in layers. Deep networks with many layers are the basis of most modern AI.
• Generative models. A large language model is trained to predict the next piece of text and can then write, summarize, translate and produce code. Systems that chain such models with tools to complete tasks are described as agentic AI.

Payment companies and card issuers use AI for fraud detection, scoring each transaction in milliseconds. Lenders use it alongside the traditional credit score to assess borrowers. Compliance teams use it to screen transactions for money laundering and to watch trading activity for signs of abuse, a task known as market abuse surveillance.

In markets, AI is applied at several points. Quantitative funds train models to forecast returns, volatility or trading volume from prices and other data. Natural language processing turns news, filings and earnings calls into numbers, for example through sentiment analysis. Execution desks use learning methods to decide how to split a large order. A robo-advisor builds and rebalances portfolios for retail clients, though most rely on standard portfolio rules more than on learning systems. AI extends algorithmic trading; it does not replace the need for a sound hypothesis, clean data and risk limits.

Financial data is a hard setting for AI. The amount of history is small compared with the data used to train image or language models, the signal is weak, and market behavior changes as participants adapt. A flexible model can fit noise in past prices and then fail in live trading, the problem of overfitting.

Other concerns are widely discussed. Complex models are hard to explain, which matters when a customer is refused credit or a regulator asks why a trade was made. Models trained on biased data can reproduce the bias. Language models can state false information with confidence. If many firms rely on similar models and data, their trades may move together and amplify market stress. Dependence on a few technology providers adds concentration risk.

Existing financial rules apply to AI regardless of the technology: a firm remains responsible for its decisions, its records and its treatment of customers. Supervisors expect model risk management, which means independent validation, monitoring of performance, and clear accountability. The European Union's AI Act, adopted in 2024, sets rules by level of risk and classes the assessment of a person's creditworthiness as a high-risk use with added duties. Securities regulators in several countries have also set requirements for testing and controlling automated trading systems, which cover those driven by AI.