Home / Wiki / G / Generative AI

Generative AI

Technology and VendorsUpdated 16 Nov 2026

Generative AI is a class of artificial intelligence systems that produce new content, such as text, images, audio, video and computer code, in response to a request. Earlier systems mostly analysed existing data: they sorted, scored or predicted. A generative system learns the patterns in a very large body of examples and uses them to compose output that did not exist before. Businesses use the technology to draft documents, answer customer questions, write software and summarise information.

The field came to wide public attention at the end of November 2022, when OpenAI released the chatbot ChatGPT, which reached an estimated 100 million users within about two months. Since then generative AI has become a major area of corporate spending and one of the most discussed themes in equity markets.

01

How it works

Generative systems are built with deep learning, a branch of machine learning that trains an artificial neural network with many layers. Two designs dominate.

  • Language models. A large language model is trained to predict the next fragment of text given what came before. Most are based on the transformer design, introduced by researchers at Google in 2017. The same approach has been extended to code, images, audio and video.
  • Diffusion models. Most image and video generators learn to turn random noise into a picture step by step, guided by a text description.

Development has two phases. In training, the model's internal settings are adjusted over weeks or months using enormous datasets. In use, known as inference, the finished model answers requests. Models that can be adapted to many tasks are called foundation models. Because output is produced from statistical patterns and not looked up from a store of facts, a model can state false information fluently, an error often called hallucination.

02

Economics

Training and running large models requires specialised chips, above all the graphics processing unit, housed in an AI data center with heavy demands for electricity and cooling. A small number of very large technology companies, each known as a hyperscaler, account for most of the spending, and their capital expenditure on AI infrastructure has risen sharply since 2023.

Developers earn revenue in several ways: consumer subscriptions, fees charged to businesses according to the volume of text processed, and features built into existing software and cloud services. The cost of producing a given quality of output has fallen quickly as chips and models improve. Total spending has nevertheless grown, since cheaper output brings more use, a pattern compared to the Jevons paradox.

03

Uses in financial services

  • Customer service. Banks, insurers and brokers use chat assistants to answer questions and guide customers through processes.
  • Document work. Staff use models to summarise research, contracts, regulatory filings and earnings calls, and to draft reports and correspondence.
  • Software. Code assistants help developers write and review programs and modernise old systems.
  • Compliance and risk. Firms apply models to reviewing communications and alerts, and to fraud detection.
  • Advice. Some wealth managers use models to support human advisers or as part of a robo-advisor service.

Regulated firms generally keep a person responsible for decisions, an arrangement known as human-in-the-loop, because of the risk of confident but wrong output.

04

Risks and regulation

  • Accuracy. Fabricated facts, figures and citations make unchecked output unsuitable for financial or legal decisions.
  • Fraud. The same tools produce convincing fake voices, images and messages, which criminals use for impersonation and scams.
  • Intellectual property. Authors, artists and publishers have brought lawsuits claiming that training on their work without permission infringes copyright. The law is unsettled.
  • Data protection and bias. Models may reveal confidential information supplied to them or reproduce unfair patterns present in their training data.
  • Concentration. Dependence on a few model developers, chip makers and cloud providers creates operational and competitive concerns.

The European Union's AI Act, adopted in 2024, sets obligations for providers of general-purpose AI models, phased in over several years. Elsewhere, regulators have mostly applied existing rules on consumer protection, data and model risk, supplemented by voluntary standards.

05

Generative AI as an investment theme

Investors approach the theme through chip makers, cloud providers, data centre and power suppliers, and software companies that use the technology. The scale of spending has prompted debate about whether revenue will grow fast enough to justify it, and comparisons with past periods of heavy investment in new technology, described under AI infrastructure cycle. Estimates of the effect on productivity vary widely, and the evidence so far is limited. Share prices linked to the theme have been volatile, and the companies that build a technology are not always those that profit most from it.

Last updated 16 Nov 2026