Definition of generative AI: what it is, how it works and real examples

"Generative AI" is on every product page, yet most definitions either say something vague like "AI that creates content" and stop, or jump straight into neural network jargon. Neither helps you judge what it can do, how it differs from the AI that has been filtering your spam for years, or when you should trust it.

This guide gives you a one-sentence definition checked against the official ones, a plain explanation of how generative AI works, a short Python demo we ran that shows the generative-versus-predictive difference in a few lines, examples by type of output, the main risks, and the rules that apply as of October 2026.

What is generative AI? The definition

Generative AI is artificial intelligence that creates new content, such as text, images, code, audio or video, by learning the patterns in large amounts of existing data and producing new output that follows those patterns. You give it a prompt, and it returns something that did not exist before: a draft email, a picture, a function, a voiceover.

The official definitions agree on the core idea and differ mainly in emphasis. Comparing them is the quickest way to see what matters:

The NIST wording is the most precise, and it is worth unpacking. "Emulate the structure and characteristics of input data" means the model learns what its training data looks like: how sentences are built, how faces are lit, how Python functions are laid out. "Derived synthetic content" means the output is newly generated, not retrieved from a database, but it is derived from those learned patterns. Both halves explain generative AI's strengths and its failures, as you will see below. The source documents are IBM's explainer (opens in a new tab), the AWS overview (opens in a new tab) and NIST AI 600-1 (opens in a new tab).

What is the main goal of generative AI?

The goal is to produce new content that is plausible and useful: it should fit the patterns of the training data and match what the prompt asked for. That is a different goal from most earlier AI. IBM sums up the contrast (opens in a new tab) this way: generative AI creates novel content, whereas predictive AI forecasts future events and outcomes.

Notice what the goal does not include: being true. A generative model is trained to produce output that looks like its data, and looking right and being right usually overlap but are not the same thing. Keep that in mind; it is the root of the biggest risk covered later.

How does generative AI work?

The details vary by model type, but almost every modern generative AI system goes through the same stages. IBM describes them as training, tuning, and generation with ongoing evaluation and retuning. In plain terms:

1. Collect a huge training set. Text from books and the web, licensed images, code repositories, audio. Generative models need far more data than most predictive models; IBM notes that predictive AI can work from smaller, more targeted datasets.
2. Pre-train a foundation model. The model repeatedly tries to predict missing pieces of unlabeled data, such as the next word in a sentence, and adjusts its internal numbers (parameters) each time it is wrong. AWS describes the result, a foundation model, as an ML model trained on a broad spectrum of generalized and unlabeled data. Our guide to what an AI model is explains parameters and training with a runnable example.
3. Tune it for people. The raw model is adapted with fine-tuning on labeled examples and with reinforcement learning from human feedback, so it follows instructions and behaves the way its developers intend.
4. Generate from a prompt. A text model writes its answer one token at a time, each time picking a likely next piece given everything so far. An image diffusion model starts from random noise and removes it step by step until a picture matching the prompt appears.
5. Ground and evaluate. Developers test outputs and keep refining the model. Many products also add retrieval-augmented generation (RAG), which feeds the model current documents at question time so answers rely on real sources instead of memory alone.

If you want the text side in more depth, including tokens, attention and a next-token demo, read what does LLM stand for. Large language models are the best-known kind of generative AI model, but not the only one.

Generative AI vs AI: see the difference in about 30 lines of Python

AI is the umbrella term for systems that do tasks normally needing human intelligence. Machine learning is the part of AI that learns from data instead of following hand-written rules. Generative AI is a branch of machine learning. The clearest way to see what makes it different is to train both kinds of model on the same data.

AWS describes the distinction well: traditional machine learning models were discriminative, focused on classifying data points, while generative models learn how features of the data relate to each other so they can produce new examples that resemble the training data. The script below does exactly that with made-up fruit weights. It uses only Python's standard library. We ran it with Python 3.13.18 on October 31, 2026.

import random
from statistics import mean, stdev

# Made-up training data: fruit weight in grams, labeled by type.
data = {
    "apple":  [150, 162, 171, 158, 166, 175, 149, 160],
    "melon":  [980, 1150, 1040, 1210, 1095, 990, 1120, 1060],
}

# Discriminative model: learn only a boundary that separates the labels.
boundary = (max(data["apple"]) + min(data["melon"])) / 2
def classify(grams):
    return "apple" if grams < boundary else "melon"

# Generative model: learn what each kind of fruit looks like
# (here, just the mean and spread of its weight).
profile = {fruit: (mean(w), stdev(w)) for fruit, w in data.items()}
def generate(fruit, n, rng):
    mu, sigma = profile[fruit]
    return [round(rng.gauss(mu, sigma)) for _ in range(n)]

print(f"discriminative boundary: {boundary:.1f} g")
print("classify 300 g ->", classify(300))
for fruit, (mu, sigma) in profile.items():
    print(f"generative profile {fruit}: mean {mu:.1f} g, spread {sigma:.1f} g")

rng = random.Random(3)  # fixed seed so you get the same output
print("new 'apples' it made up:", generate("apple", 5, rng))
print("new 'melons' it made up:", generate("melon", 5, rng))

Running python3 gen_vs_disc.py printed exactly this:

discriminative boundary: 577.5 g
classify 300 g -> apple
generative profile apple: mean 161.4 g, spread 9.2 g
generative profile melon: mean 1080.6 g, spread 79.0 g
new 'apples' it made up: [162, 173, 153, 171, 159]
new 'melons' it made up: [1060, 1231, 1093, 1077, 1138]

Toy as it is, this shows the whole difference:

• The discriminative model only learned a line. Anything under 577.5 g is an apple. It can label, but it cannot make anything, and it confidently calls a 300 g object an apple even though nothing in its data weighed that much.
• The generative model learned what the data looks like. It stored a profile of each fruit and can sample new examples from it. The samples are fresh draws, not values looked up from the list: some happen to match training weights, others such as 153 g and 173 g do not, and one invented melon, 1,231 g, is heavier than any melon it saw. All of them are plausible.
• Plausible is not the same as real. Those made-up weights describe fruit that does not exist. Scale the idea up to billions of parameters trained on text, and you get a model that writes fluent paragraphs, including fluent paragraphs that are wrong.

Real generative AI models learn vastly richer profiles than a mean and a spread, but the principle is the same: learn the distribution of the data, then sample new content from it.

One more distinction you will hear: agentic AI. That is not a different kind of model but a generative model placed in a loop with tools and permissions so it can take actions toward a goal. We cover that comparison separately in agentic AI vs generative AI.

Which technique is commonly used in generative AI?

Today's generative AI is built on deep learning: large neural networks trained on huge datasets. Within deep learning, four model families did most of the work. IBM gives the timeline, and the original papers are linked where we checked them:

The GAN paper is a neat illustration of the generative-versus-discriminative split from the demo above: it trains a generative model that captures the data distribution against a discriminative model that estimates whether a sample came from the real data. IBM notes that transformers underpin most of the leading generative AI tools today, which is why "generative AI model" and "large language model" are often used as if they meant the same thing. They overlap heavily, but image and video generators are generative AI too.

Examples of generative AI

Generative AI examples are easiest to sort by what they produce. The products change quickly, so this table lists the type of output first and well-known examples second.

Usage at work has grown fast. In McKinsey's global state of AI survey, the share of respondents reporting gen AI use (opens in a new tab) in their organizations rose from 33% in 2023 to 71% in 2024 and 79% in 2025. The 2025 survey ran from June 25 to July 29, 2025, with 1,993 participants. The chart's title sums up McKinsey's caveat: AI at work, but not at scale.

Schools are a special case, with real benefits and real concerns. Our guide to AI in education covers that debate with evidence on both sides.

When generative AI is a good fit, and when it is not

Because generative AI aims for plausible output, it works best where a person can quickly judge the result and where a first draft has value even if it needs edits. A quick checklist:

• Good fit: first drafts, rewording and summarizing material you can check, brainstorming options, explaining unfamiliar code or concepts, turning notes into structure, creating variations of something you already approved.
• Use with checks: factual research, numbers, legal or medical wording, and anything published under your name. Ask for sources, ground the model in your own documents with RAG, and verify before you use the output.
• Poor fit: decisions that need an exact, auditable answer every time, such as calculating a payroll or approving a loan. A rule-based system or a predictive model you can test is usually safer and cheaper.

How you ask matters too. Clear instructions, examples and context improve output from any model; our Claude prompt engineering guide shows techniques that work across assistants.

The risks of generative AI

NIST's Generative AI Profile lists 12 risks that are unique to generative AI or made worse by it. Four matter to almost every user:

• Confabulation, better known as hallucination. NIST defines it as the production of confidently stated but erroneous or false content by which users may be misled or deceived. This is the toy demo's made-up melon, at scale.
• Information integrity. Generative AI lowers the cost of producing content that may not separate fact from opinion or fiction, and that could be used for large-scale disinformation campaigns.
• Intellectual property. It eases the production or replication of copyrighted, trademarked or licensed content without authorization, and can expose trade secrets.
• Data privacy and harmful bias. NIST lists both, and AWS notes that outputs can reflect biases in the training data. AWS also notes that generative models are often considered black boxes because their specific outputs are hard to explain.

The NIST list also covers environmental impacts and dangerous, violent or hateful content, among others. None of these mean you should avoid generative AI. They mean you should match the level of checking to the stakes of the task.

Generative AI rules as of October 2026

Two developments are worth knowing if you publish or sell anything made with generative AI.

EU: labeling AI-generated content

Article 50 of the EU AI Act applies from 2 August 2026 (opens in a new tab), according to the European Commission. Providers of generative AI systems must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Systems already on the market before 2 August 2026 have until 2 December 2026 for that marking duty. Separately, anyone deploying a system to create deepfakes must disclose them clearly to people at first exposure, with a lighter duty for evidently artistic, satirical or fictional work. Content generated before 2 August 2026 does not have to be labeled retroactively.

US: copyright in AI output

On January 29, 2025, the US Copyright Office released Part 2 of its AI report (opens in a new tab), on copyrightability. It concluded that AI-generated outputs can be protected only where a human author has determined sufficient expressive elements, for example through creative arrangement or modification, and that the mere provision of prompts is not enough. If you need to own what you make, your own creative contribution has to be real and documentable.

The bottom line

The definition of generative AI comes down to one idea: a model learns the structure of a large body of data, then creates new content that follows it when you prompt it. That is what separates it from predictive AI, which labels and forecasts, and it is also why its output can be fluent and wrong at the same time. Use it where a draft has value and you can check the result, ground it in real sources when facts matter, and keep a person responsible for what goes out. For more plain-English explainers on models, agents and tools, browse our AI guides.