Generative AI

What it means

Generative AI is software that makes something new (a paragraph, a picture, a voice recording) after learning from a very large collection of earlier examples. The United States government once defined it as AI that copies "the structure and characteristics of input data in order to generate derived synthetic content," including "images, videos, audio, text, and other digital content."¹

Here is an everyday example. A manager asks a chat assistant to draft a thank-you note to a retiring colleague. The software has never seen that note, because the note did not exist before. It writes one by drawing on patterns from the huge amount of writing it learned from. The draft is new, and whether it is good is still the manager's call.

Where the word comes from

We could not find a reliable record of who first used the exact phrase "generative AI," so we make no claim about it. The idea grew out of research on "generative models," meaning programs built to produce new examples rather than only sort existing ones. A 2014 paper by Ian Goodfellow and colleagues, which proposed "a new framework for estimating generative models via an adversarial process," is one of the best-known milestones.² IBM Research gives a short history. Variational autoencoders, "introduced in 2013," were "the first deep-learning models to be widely used for generating realistic images and speech." Transformers were "introduced by Google in 2017," and GPT-3 "was the largest language model of its kind when OpenAI released it in 2020."³ The word started as a research term. It is now a label for a whole category of products.

How others define it

• US Executive Order 14110 (2023): "The term 'generative AI' means the class of AI models that emulate the structure and characteristics of input data in order to generate derived synthetic content. This can include images, videos, audio, text, and other digital content."¹ The order was rescinded in January 2025,⁴ but the US standards agency NIST built a Generative AI Profile of its risk framework "pursuant to" that order.⁵
• IBM Research (2023): "Generative AI refers to deep-learning models that can generate high-quality text, images, and other content based on the data they were trained on."⁶
• EU AI Act, Recital 99: "Large generative AI models are a typical example for a general-purpose AI model," because "they allow for flexible generation of content, such as in the form of text, audio, images or video."⁷

Where they agree and differ

All three agree on the core idea: the system learns from existing data and then makes new content. They also list the same kinds of output, which are text, images, audio, and video.

They differ in emphasis. The US definition is the broadest and covers any "class of AI models" that does this. IBM narrows it to "deep-learning models" (a kind of software loosely inspired by the brain that learns from examples) and adds "high-quality," which is a judgment and not a test. The EU text treats generative AI as an example of a bigger legal category, general-purpose AI, and stresses how flexible it is.

There is a fair objection to all three. Because the definitions are so broad, the label can cover almost any modern AI product, which makes it a weak guide to risk. A tool that suggests email subject lines and a tool that writes medical summaries are both "generative AI." We agree that the label alone tells a leader little. What matters is what the content is used for and who checks it.

Generative AI is also not the same as an [AI agent](/definitions/ai-agent). Generative AI makes content. An agent uses that ability to take actions, such as sending an email or changing a record, often through [tool use](/definitions/tool-use).

In the news

On October 16, 2026, the European Commission published "the most comprehensive update" of its ERA Living Guidelines on the responsible use of generative AI in research.⁸ The update adds guidance on agentic AI. The Commission notes that such tools "are becoming more autonomous and capable of handling increasingly complex tasks," "making questions of oversight, accountability and reliability all the more important."⁸ It also says that "Research integrity principles, such as accountability, transparency and responsibility, are still the main reference."⁸ We are confident in this account because it comes from the Commission's own page. The point for leaders is that a European rule-writer is now treating generative AI and agents as one continuing question of oversight.

What it means for your organization

Generative AI produces drafts, and the people who use those drafts are still responsible for them. The Commission's guidelines make the same point when they say integrity principles "are still the main reference."⁸ Once generative AI is built into agents that act, it matters more that [a person stays in the middle](/definitions/a-person-in-the-middle) of each chain of command, whether that is a [human in the loop](/definitions/human-in-the-loop) or clear [guardrails](/definitions/guardrails).

One question to ask your team: Where does content made by generative AI leave our building, whether to customers, regulators, or the public, without a named person reading it first?