What is generative AI? How it works, examples, risks and rules

What makes AI generative?

Predictive AI, sometimes called discriminative AI, looks at an input and returns a label or a number: spam or not spam, a fraud score, next month's demand. Generative AI returns content. The definition in NIST's Generative AI Profile (NIST AI 600-1, July 2024) says the same thing more formally: models that emulate the structure and characteristics of their input data to generate derived synthetic content, including images, video, audio and text.

Two properties explain most of what follows:

• The output is sampled, not retrieved. The model generates what is likely given its training and your prompt. The same prompt can give different answers, and a confident answer can be invented.
• The largest models are general purpose. One model drafts an email, writes SQL and summarizes a contract, steered only by the instructions in the prompt. The EU AI Act calls these general-purpose AI models: models that display significant generality and can competently perform a wide range of distinct tasks. The European Commission's indicative test is a model trained with more than 10²³ floating-point operations (FLOP) that can generate language, images from text or video from text.

Generative AI examples, grouped by what they produce:

How does generative AI work?

Every generative model is a neural network: a very large set of adjustable numbers, called parameters or weights, that training tunes until the network's output resembles its training data. The families differ in what the network learns to predict. How such a network learns in the first place, from weights and backpropagation to embeddings, is explained in deep learning and NLP explained.

Large language models predict the next token

A large language model (LLM) reads and writes tokens, which are words or pieces of words. For Google's Gemini models, a token is about four characters, and 100 tokens are about 60 to 80 English words. Given the tokens so far, the model scores every possible next token, picks one, appends it and repeats until the answer is complete. That is the mechanism behind GPT-3, which OpenAI researchers described in May 2020 as an autoregressive language model with 175 billion parameters.

Two consequences matter in practice. The model commits to each token before it knows how the sentence will end, so an early mistake is carried through to a confident conclusion. And because the next token is sampled rather than always the single most likely one, the same question can produce different answers; settings such as temperature control how much variety you get.

The network design underneath is the transformer, introduced in the 2017 paper Attention Is All You Need. Its attention mechanism lets each token weigh every other token in view, which is how a model keeps track of what "it" refers to three paragraphs back. Meta, for example, describes Llama 3.1 as a standard decoder-only transformer, and extended its context window, the amount of text the model can see at once, to 128K tokens in July 2024.

Diffusion models turn noise into images

Stable Diffusion is a diffusion model. In training, the model sees images with increasing amounts of random noise added and learns to predict and remove that noise, the method set out in Denoising Diffusion Probabilistic Models in 2020. Generation runs the process backwards. The High Court in London described it plainly in its 2025 judgment on Stable Diffusion: generation starts from a random noise image, and the trained network, conditioned by the user's prompt, removes the noise step by step until the image is consistent with the prompt. The same judgment records that the model does not store its training data; what training leaves behind is the learned weights.

The judgment also identifies Stable Diffusion as a latent diffusion model, from the latent diffusion paper first published in December 2021. Denoising runs on a compressed representation of the image instead of every pixel, which cuts the computing cost, and cross-attention layers let a text prompt steer each step. Video generation has followed: Google's Veo 3, announced on May 20, 2025, was the first Veo model to generate sound along with the picture, from street noise to dialogue.

Multimodal models handle several kinds of data in one network

A multimodal model takes in, and often produces, more than one kind of data. Before GPT-4o, announced on May 13, 2024, ChatGPT's voice mode chained three models (speech to text, a text model, then text to speech) and lost tone, multiple speakers and background sound along the way. OpenAI trained GPT-4o end to end across text, vision and audio, so one network processes all of them: it accepts any combination of text, audio, image and video, and outputs text, audio and images. Google's Gemini 3 announcement credits the first Gemini generation with native multimodality and long context, the base the later generations built on.

How generative AI models are trained

Training happens in stages. Pretraining teaches the model language, code or images; the later stages teach it to follow instructions and to behave the way people want.

1. Pretraining. The model learns from a huge unlabelled dataset by predicting the next token, or by removing noise for images. The data is its own answer key, which is why this is called self-supervised learning. Meta trained Llama 3.1 405B on more than 15 trillion tokens, using more than 16,000 H100 GPUs. The European Commission notes that training a model at the AI Act's 10²⁵ FLOP threshold is estimated to cost tens of millions of euros, which is why organizations start from a pretrained model rather than training their own.
2. Instruction tuning. People write example prompts with good answers, and the pretrained model is fine-tuned on them (supervised fine-tuning, or SFT). This is what makes it answer a question instead of continuing your text.
3. Reinforcement learning from human feedback (RLHF). People rank several answers to the same prompt, a reward model learns their preferences, and the language model is tuned to score well against it. In OpenAI's InstructGPT paper (March 2022), people preferred the answers of a 1.3-billion-parameter model trained this way to those of the 175-billion-parameter GPT-3. ChatGPT was trained with the same method. Meta's post-training for Llama 3.1 used several rounds of supervised fine-tuning and direct preference optimization (DPO), another way to learn from ranked answers.
4. Reasoning training. With OpenAI o1 in September 2024, large-scale reinforcement learning began teaching models to work through a problem in a chain of thought before answering. OpenAI reported that performance kept improving with more training and with more time spent thinking. DeepSeek-R1, released on January 20, 2025 under the MIT licence, brought the approach to an open-weights model.

One consequence of this pipeline is easy to forget: a model knows nothing that happened after its training data was collected. The first ChatGPT, for example, was fine-tuned from a GPT-3.5 model that finished training in early 2022. Current facts, and your own organization's facts, have to be supplied in the prompt, which is the job of retrieval (below).

Generative AI timeline: key milestones from 2017 to 2026

Each date below comes from the publisher's own paper or announcement. Model names change every few months; the mechanisms above change much more slowly.

Read down the right-hand column and the direction is clear: from writing text, to handling images, sound and video, to reasoning, and by 2026 to operating software on a user's behalf. Open-weights releases (Stable Diffusion, Llama 3.1, DeepSeek-R1) appear throughout; they are the option for organizations that need to run a model on their own infrastructure. Our Gemini explainer follows one model family in more detail, and our post on GPT-4 Turbo follows one OpenAI model from its 2023 launch to its scheduled shutdown in October 2026.

What businesses use generative AI for

The business uses that hold up share one shape: the model drafts or finds, and a person or a rule decides what happens next.

Drafting and summarizing

Summarizing a document you supply is safer than asking a model what it knows: the source is in front of the model, and a reviewer can check the summary against it. Drafting from the model's memory is where invented facts slip in, so the first draft is where the model's job ends and an editor's begins. If the drafts are web pages, read how search engines treat them in AI-generated content and SEO. For the drafting tools built into Gmail, Copilot and ChatGPT, and how to use them safely at work, see what Draft AI is.

Coding help

Coding assistants suggest completions, explain code, write tests and, as agents, make changes across many files. The evidence on productivity depends on the task. In a 2023 controlled experiment, developers with GitHub Copilot implemented an HTTP server in JavaScript 55.8% faster than developers without it. In a 2025 randomized trial by METR, 16 experienced open-source developers working on mature projects they knew well took 19% longer with early-2025 AI tools, even though they estimated afterwards that the tools had saved them 20%. The two studies measured very different work, which is the lesson: measure on your own tasks before you assume a number. Our guide to the future role of software engineers sets out every trial and what it means for the job.

Customer support

A study of 5,179 customer support agents (NBER working paper, 2023) found that a generative AI assistant raised the number of issues resolved per hour by 14% on average and by 34% for novice and low-skilled agents, with minimal effect on the most experienced. The design is the point: the assistant suggests, and the agent decides what the customer receives.

Search over internal documents with retrieval

Retrieval-augmented generation (RAG), described by Lewis et al. in 2020, is the pattern behind "chat with our documents" tools. The question is first used to search your own content, the best passages are placed in the prompt, and the model answers from them, citing the passage it used.

Microsoft's Azure AI Search documentation lists the hard parts: understanding vague questions, reaching content spread across SharePoint, databases and storage, fitting results into a limited token budget, and making sure users and agents retrieve only content they are authorized to see. Retrieval gives the model the right facts; it does not guarantee the model uses them correctly, and it opens a new attack path (prompt injection, below). For questions over structured data, the equivalent is text-to-SQL, covered in generative AI for databases.

How widely is generative AI used?

Adoption figures depend on who is asked.

• Surveys of managers. McKinsey's Global Survey on AI found in mid-April 2023 that one-third of respondents' organizations used generative AI regularly in at least one business function. In February and March 2024 the share was 65%.
• The latest survey. McKinsey's 2026 edition, fielded from May 4 to June 8, 2026 with 1,719 participants in 97 countries, reports its headline figures for AI as a whole. Nearly nine in ten respondents report regular AI use in at least one function and 44% say AI is scaling across their enterprise, but only 37% attribute any EBIT impact to it, about the same share as a year earlier.
• Official statistics. Eurostat found that in 2025, 19.95% of EU enterprises with 10 or more employees used any AI technology: 8.76% used AI that generates written or spoken language or code, and 9.55% used AI that generates pictures, video or audio. Among large enterprises, 55.03% used AI.

The gap is not a contradiction. A survey of managers measures something different from a statistical survey of all enterprises, and 36% of McKinsey's 2026 respondents work for organizations with more than $1 billion in annual revenue. Read together: most large organizations now use AI, most smaller ones do not yet, and in McKinsey's data the financial impact lags behind the use.

The risks of generative AI

NIST's Generative AI Profile lists 12 risks that generative AI creates or makes worse. The sections below cover the four that any business deployment has to handle (confabulation, intellectual property, data privacy and information security, which includes prompt injection), plus one NIST does not list: cost.

Hallucination

Models state false things confidently. NIST calls this confabulation, and the first limitation OpenAI listed when it launched ChatGPT was answers that sound plausible but are incorrect. The 2025 paper Why Language Models Hallucinate argues that the causes are statistical: training and evaluation reward guessing over admitting uncertainty. The consequences land on whoever deploys the system. In Moffatt v. Air Canada (February 2024), British Columbia's Civil Resolution Tribunal rejected the airline's argument that it could not be held liable for its chatbot, found that it was responsible for all the information on its website, and ordered it to pay damages to a customer the chatbot had misinformed about bereavement fares. Our post on AI limitations in understanding explains why fluent models fail this way.

Copyright disputes still in court

Two questions are open: whether training on copyrighted work without a licence is lawful, and who owns what a model produces.

• United States. The New York Times v. Microsoft and OpenAI, filed on December 27, 2023 in the Southern District of New York, was still active in September 2026. The U.S. Copyright Office released a pre-publication version of its report on generative AI training on May 9, 2025; the courts, not the Office, will decide the cases.
• United Kingdom. In Getty Images v. Stability AI (November 4, 2025), the High Court held that a model which does not store or reproduce the copyright works is not an "infringing copy". Getty won only a trade mark point about watermarks that the judge called historic and extremely limited, and it had abandoned its training claim because there was no evidence the training took place in the UK.
• Outputs. The Copyright Office concluded in January 2025 that AI output is protected by copyright only where a human author determined sufficient expressive elements. Prompts alone are not enough.

Privacy

Privacy risk runs in both directions. What staff paste into a prompt goes to the provider, and whether it is kept or used for training depends on that provider's terms for your plan. What a model learned can also leak: researchers extracted hundreds of verbatim sequences from GPT-2's training data, including names, phone numbers and email addresses, simply by querying it (USENIX Security 2021). In the EU, the European Data Protection Board said in December 2024 that a model trained on personal data counts as anonymous only if identifying people, or extracting their data through queries, is very unlikely, assessed case by case.

Prompt injection

Prompt injection is to language models what SQL injection is to databases: data the system ends up treating as instructions. It is LLM01 in OWASP's 2025 Top 10 for LLM applications. A direct injection comes from the user's own prompt; an indirect one hides in content the model reads, such as a web page, an email or a retrieved document, and it does not have to be visible to a human. OWASP notes that RAG and fine-tuning do not fully mitigate it.

Cost

Generative AI is billed by use, and the bill grows in ways a licence fee does not. APIs charge per million tokens, and output costs more than input. On Google's Gemini API price list as of September 2026, Gemini 3.1 Pro Preview costs $2.00 per million input tokens and $12.00 per million output tokens for prompts up to 200,000 tokens, and the output price includes the model's thinking tokens. Agents cost more again, because every intermediate step in their loop is billed as tokens. In McKinsey's 2026 survey, about 20% of respondents said AI operating costs, including token costs, constrained their use of AI.

Generative AI regulation: the EU AI Act and other jurisdictions

The EU AI Act, Regulation (EU) 2024/1689, entered into force on August 1, 2024 and applies in stages. The dates below come from the Act and the Commission's timeline, which includes the extensions that followed the political agreement on the AI Omnibus simplification proposal.

According to the Commission's questions and answers, a provider of a general-purpose model must draw up technical documentation for downstream providers and authorities, put a copyright policy in place and publish a summary of the content used for training. The Act presumes that a model trained with more than 10²⁵ FLOP has high-impact capabilities, which makes it a model with systemic risk; its provider must also notify the Commission, evaluate and mitigate the risks, report serious incidents and secure the model. The General-Purpose AI Code of Practice, published on July 10, 2025, is the voluntary route to showing compliance. Under Article 101 of the Act, which applies from August 2, 2026, the Commission can fine providers of these models up to 3% of annual worldwide turnover or €15 million, whichever is higher.

If you use these models rather than build them, most of those duties sit with the provider. The rules most likely to reach a business that deploys generative AI are the transparency rules: since August 2, 2026, chatbots must tell people they are dealing with AI, deepfakes must be labelled, and AI-generated content must carry machine-readable marks. Providers are building the marks in; Anthropic, for example, says it watermarks text from its models released after August 2, 2026.

Elsewhere, briefly:

• China has regulated public generative AI services since August 15, 2023, under the Interim Measures for the Management of Generative AI Services issued by the Cyberspace Administration of China and six other agencies.
• United States. California's SB 53, the Transparency in Frontier Artificial Intelligence Act, signed on September 29, 2025, requires large frontier developers to publish a framework describing how they apply national and international standards and industry best practices.
• Canada. The proposed Artificial Intelligence and Data Act, part of Bill C-27, was still in committee when that session of Parliament ended on January 6, 2025, so it did not become law. Personal data you send to a model is governed by existing privacy law.

How to start using generative AI in your organization

Start with a task small enough to measure. A broad "assistant for everyone" is hard to evaluate and harder to fix.

1. Pick one narrow, frequent task where a draft or an answer saves time and a person already checks the result: summarizing support tickets, drafting replies from a knowledge base, pulling fields out of invoices.
2. Decide what data may go to which model, under which provider terms, before anyone pastes a document into a chat window.
3. Build an evaluation set from real cases with known good answers, including the awkward ones. Anthropic's guidance on evaluations makes the same point: test cases should mirror the real task, include edge cases and be graded automatically where possible.
4. Prototype against that set, with retrieval over approved sources if the task needs your facts. Record accuracy and cost per case.
5. Keep a person in the loop for anything that leaves the organization or changes a record, and log what the model saw and produced.
6. Re-run the evaluation before every model, prompt or source change. The timeline above shows how often models are replaced, and a new version can be better on average and worse on your task.

If you want help choosing and testing a first use case, our AI and automation service starts with one real workflow, reviewed with the people who run it, before anything is built, and every model, prompt or source change must answer the same set of real questions before it ships. For building AI into a product rather than a workflow, see intelligent apps.