What Is Generative AI?

Generative AI is artificial intelligence that creates new content, such as text, images, audio, video or code, in response to a request called a prompt. Much earlier AI was built to analyze existing information, for example classifying an email as spam. Generative AI produces something new instead. It learns patterns from huge amounts of example data and uses them to generate original output. ChatGPT and AI image generators are common examples.

What generative AI means

The "generative" in generative AI means it generates: it produces new material rather than only sorting, scoring or labeling what already exists.

The U.S. National Institute of Standards and Technology defines generative AI as "the class of AI models that emulate the structure and characteristics of input data in order to generate derived synthetic content." In other words, a generative model studies many examples of something, such as sentences, photos or songs, and learns to produce new content with similar characteristics.

The difference from other AI is easiest to see side by side.

Both kinds often run on similar technology, and many modern products combine them. The LATHIC Glossary has the short definition of generative AI.

How generative AI works

Generative AI works in two broad phases: learning from data, then generating from a prompt.

Phase 1: Learning patterns from data

A generative model is trained on a very large collection of examples. A language model learns from text. An image model learns from images, usually paired with descriptions. During training, the model adjusts millions or billions of internal settings until it captures the patterns in that data: how words tend to follow each other, what a "sunset over the ocean" typically looks like, how a certain voice sounds.

Large, broadly trained models are often called foundation models, because they can be adapted for many different tasks. After this initial training, developers usually refine them further. For chat assistants, that includes training on human feedback so the model follows instructions and responds helpfully, an approach OpenAI described in its 2022 InstructGPT research.

Phase 2: Generating from a prompt

When you give a generative AI tool a prompt, the model uses the patterns it learned to produce new output. Different types of content are generated in different ways.

Text. Large language models generate text one small piece at a time. OpenAI explains that its models use what they learned about how words typically appear together to predict the next most likely word when generating a response. Repeating that prediction produces sentences, paragraphs and entire documents. What Is a Large Language Model? explains this in more depth.

Images and video. Many image and video generators use diffusion models. NIST describes how they "start from a full output consisting of random noise, then iteratively refine the noise into an output matching the prompt." Each refinement step makes the image a little clearer and closer to the description.

Earlier approaches. Before diffusion models became common, many image generators used generative adversarial networks, or GANs, introduced in 2014. A GAN trains two neural networks against each other: one creates images and the other tries to tell them apart from real ones.

Models usually choose from several likely options rather than always picking the single most likely word or pixel, so the same prompt can produce different results each time. That variety is useful for creative work, and it is also why generative AI can be inconsistent.

What generative AI can create

Generative AI now works across many types of content, often called modalities.

What is generative AI used for?

People and organizations use generative AI for:

• Writing and editing: emails, reports, marketing copy, first drafts.
• Summarizing: long documents, meetings, research.
• Brainstorming: ideas, outlines, names, options.
• Customer support: answering questions and drafting responses.
• Software development: writing and reviewing code.
• Design and media: concept images, product mockups, video and audio.
• Learning: explanations, practice questions and tutoring.

Use is now widespread. Stanford's 2026 AI Index reported that generative AI is used in at least one business function at 70% of the organizations in the surveys it cites. A Pew Research Center survey conducted in February 2026 found that about half of U.S. adults use AI chatbots.

Generative AI vs AI

AI is the broad field of building computer systems that perform tasks associated with human intelligence. Generative AI is one type of AI, focused on creating new content. All generative AI is AI, but much AI, such as the systems behind fraud detection, navigation estimates or face recognition on phones, analyzes information rather than generating it.

Generative AI vs LLMs

A large language model is one kind of generative AI model, specialized in language. Generative AI is the wider category that also includes models that generate images, video, audio and more.

In practice, the terms overlap because language models power the most widely used generative AI tools, and many newer models can handle several types of content. A useful shorthand: nearly every LLM in use today is generative AI, but not every generative AI system is an LLM.

Generative AI vs agentic AI

Generative AI creates content in response to a prompt, and the next step is usually yours. Agentic AI uses generative models together with tools and planning to take actions toward a goal, such as searching, filling in forms or sending messages. Most agentic systems are built on top of generative AI. For a full comparison, see What Is Agentic AI?.

Where generative AI fails

Generative AI is useful, but it has predictable weaknesses.

It can make things up. Generative AI can produce false information that sounds convincing. NIST calls this "confabulation," the production of confidently stated but erroneous or false content, and notes it is known colloquially as "hallucinations." When OpenAI launched ChatGPT in 2022, it acknowledged the tool "sometimes writes plausible-sounding but incorrect or nonsensical answers."

It reflects its training data, including its biases. NIST lists "harmful bias or homogenization" among the risks of generative AI, including the amplification of historical, societal and systemic biases and performance differences between groups or languages.

It can be out of date. A model's built-in knowledge comes from data collected before a certain date. Unless a tool can search for current information, it may not know about recent events.

It is uneven. The International AI Safety Report 2026 notes that general-purpose AI can excel at complex tasks such as generating code or answering expert-level questions, yet struggle with seemingly simpler ones, such as counting objects in an image, reasoning about physical space or recovering from basic errors in longer workflows.

It raises intellectual property questions. NIST identifies intellectual property as a generative AI risk, including easier production of content that may copy copyrighted or trademarked material without authorization. Legal questions about training data and outputs are still being worked out in courts and legislatures.

It can be misused. The same tools that create useful images and audio can create convincing fakes, including impersonations using voice cloning.

It is not the same as understanding. Generative AI produces output based on learned patterns. It can write about a topic fluently without verifying whether what it writes is true.

For why made-up answers happen and how to reduce them, see What Is an AI Hallucination?.

How generative AI developed

Generative AI became mainstream quickly, but it built on more than a decade of research:

• 2014: Researchers introduced generative adversarial networks (GANs), an approach that later produced far more realistic AI-generated images.
• 2017: Google researchers introduced the Transformer, the architecture behind today's large language models.
• 2020: OpenAI described GPT-3, a language model with 175 billion parameters that could write convincing text from simple prompts. The same year, researchers showed high-quality image generation with diffusion models.
• 2022: OpenAI released ChatGPT on November 30, bringing generative AI to a mass audience.
• 2025 onward: Generative models expanded into video with sound, and into agentic systems that use generated plans and text to carry out tasks.

Why generative AI matters

Generative AI changed who can use AI. Earlier AI mostly worked behind the scenes, built into products by engineers. Generative AI lets anyone ask for what they want in everyday language and get a result in seconds.

It also shifts responsibility to the person using it: deciding what to ask, checking what comes back and choosing what to trust. For how generative AI fits into the wider picture of how AI systems are built and used, see How Does AI Work?.