What is generative AI? Meaning and examples (2026)

Q: What does GenAI stand for?

GenAI is short for generative artificial intelligence. It means the same as generative AI.

In short

Generative AI (GenAI) creates new content. Traditional AI classifies or predicts from options that already exist.
A language model keeps predicting the next piece of text. It calculates probability, not truth.
In 2026 generative AI increasingly works as an agent: it searches systems, uses tools and puts a proposal in front of you.
The biggest risk is invented facts and sources. Connecting the model to real sources and having a person check the result reduces that risk.
Since 2 August 2026, the EU AI Act requires providers to mark AI-generated content in a machine-readable way (Article 50).

What is generative AI?

Generative AI is artificial intelligence that creates new content. The model learns patterns from large amounts of existing data and uses them to produce something new: a text, an image, speech, a table or a piece of source code.

The model does not copy an existing sentence. It composes a new one that fits what it has learned. That is how it can write a letter, summary or analysis that exists nowhere else.

Two misunderstandings cause most of the trouble in practice:

Generative AI is not a search engine. A search engine points you to a document. Generative AI writes an answer itself. That answer can be right without any document behind it. It can also be wrong without the text giving you any hint.
Generative AI is not a decision-maker. The model delivers a proposal. Whoever adopts it is responsible for the content. See also who is liable when an AI agent causes harm.

What does generative AI mean?

"Generative" means producing: something that brings forth something new. The word comes from the Latin generare, to bring forth. Generative AI is AI that produces content, as opposed to AI that only sorts or predicts.

You will see the concept under several names:

Term	Where you see it
Generative AI	the common English term
GenAI	short for generative artificial intelligence
Foundation model	a large model trained on broad data that many applications build on
Large Language Model (LLM)	a generative model specialised in language
Generatieve AI	the Dutch term, with the same meaning

How does generative AI work?

A language model keeps predicting the next piece of text. The scale is what makes this remarkable, not the idea. It works in four steps:

Text is split into tokens. A token is a word or part of a word. "Unbelievable" might become "un", "believ" and "able".
The model calculates the most likely next token. After "The tenant handed back the keys on 1 ...", "March" and "April" are likely. "Banana" is not.
Attention decides what counts. The transformer architecture weighs, for each token, how much every earlier word matters. That is how the model still knows twenty sentences later that "she" is the tenant.
The answer grows token by token. Each chosen token goes back into the input. There is deliberately a little randomness in that choice, which is why you rarely get exactly the same answer twice.

Four-step diagram: text becomes tokens, the model calculates the most likely next token, attention weighs earlier words and the answer grows token by token. The model calculates probability, not truth.

The model does not calculate truth, it calculates probability. A fluent and convincing answer is exactly what it was trained to produce, whether or not that answer is correct.

The knowledge sits in billions of numbers: the parameters. They are formed in two phases. First the model reads vast amounts of text and learns language, which is called pre-training. Then it learns from examples of good answers and from human preferences, so that it follows instructions.

What is different in 2026. Current models differ from the chatbots of a few years ago in three ways:

They think first. Reasoning models work through a question in intermediate steps before they answer. That takes longer, but it helps on hard questions.
They are multimodal. They read not only text but also scans, tables, drawings and speech.
They use tools. A model can look something up, query a source or run a calculation. The Model Context Protocol (MCP) has become the open standard for this. In December 2025 Anthropic moved MCP to the Agentic AI Foundation under the Linux Foundation.

What is the difference between generative AI and traditional AI?

Traditional AI chooses or predicts from existing options. Generative AI makes something new. That difference also decides how you check the result.

	Traditional (predictive) AI	Generative AI
The question	"Which box does this belong in?" or "How much will it be?"	"Make something new from this"
The output	a label, a number, a probability	text, images, code, a summary
Example	spam filter, fraud detection, predictive maintenance	draft letter, case summary, meeting notes
Typical error	a case put in the wrong box	fluent and wrong at the same time
How you check it	the error rate across many cases	each answer on its own, against the source

The last row matters most. With a spam filter you accept an error rate across thousands of messages. With generative AI an average does not help you: one invented reference in one letter is already a problem.

What are LLMs, and how do they relate to AI agents?

A Large Language Model (LLM) is a generative model for language. It is trained on books, news, legislation, forums and source code. Well-known model families are GPT (OpenAI), Claude (Anthropic), Gemini (Google), Llama (Meta), Mistral (France), and DeepSeek and Qwen (China).

The terms around generative AI are not synonyms. They are layers that fit inside each other:

Artificial intelligence (AI): the whole field.
Machine learning: systems that derive their behaviour from data rather than fixed rules.
Deep learning: machine learning with neural networks of many layers.
Generative AI: deep learning that creates new content. This is where LLMs sit.
AI agents: generative AI that uses tools and carries out a task. An agent looks things up, calculates, drafts and reviews its own intermediate result.

A chatbot answers a question. An agent carries out a task. That changes the control question too. It is no longer only "is this answer right?" but also "which steps were taken and which sources were used?"

What AI agents do in practice, and how several agents work together as virtual colleagues, is covered in What are AI agents and how do they help you move forward?.

What are examples of generative AI in 2026?

The strongest applications have one thing in common: the model works with real, verified sources instead of its memory.

Type	Example from practice
Text	a draft letter, the summary of an annual report, minutes from a recording
Research	searching court decisions by area of law and court, with the full text of each ruling
Documents and images	reading a scanned deed or building drawing and comparing it with the requirements
Speech	writing a care report from what a nurse dictates
Code	having a function written and then tested
Agents	an assistant finds risks in policy documents and submits proposals for approval

In the Netherlands, where we work, it looks like this per sector:

Legal work. Finding court decisions, retrieving the full text and seeing which rulings cite each other. Since December 2025 Dutch lawyers have recommendations from their bar association (NOvA) on AI use. Read what the NOvA recommendations mean for your firm. How that research works with an AI assistant is shown on our page on legal research with AI (in Dutch).
Public administration. For a freedom of information request, AI marks the passages with personal data and proposes redactions. The case officer decides.
Construction. Checking a building plan against the Dutch building regulations and quality assurance act, and listing where the substantiation is missing.
Care. Writing a report from what a care professional dictates, linked to the client file.
Internal knowledge. Staff ask for a work instruction in plain language and get an answer with a link to the document.

How reliable is generative AI?

Generative AI is as reliable as the sources you connect it to and the checks you organise. Without sources, a model sometimes invents facts, and they look just as convincing as real ones.

Where it goes wrong

Hallucination. The model invents something that sounds logical but does not exist. Ask an assistant without a source connection for a Dutch Supreme Court ruling, and you may get a neatly formatted case number that points nowhere. See why AI invents Dutch case law.
Citations that are wrong. Without a connection to a source system, a model cannot give a real reference. Ask for one anyway and it will make one up.
Outdated knowledge. A model knows the world up to its training date. Last month's change in the law is not in it, unless the model can look it up live.
Bias. What is common in the training data comes back in the output. In texts about people, that is a real risk.
Accepting too quickly. A fluent answer does not invite scrutiny. That is why checking has to be organised.

What helps

Measure	What it does
RAG (retrieval-augmented generation)	gives the model passages from an indexed knowledge base, so the answer rests on documents
Live connection	lets the model query the source at the moment of the question, for example via MCP
Clickable citations	make checking possible in one click
Four eyes	a subject expert reviews the result before it leaves the building
Second model as reviewer	another model tests whether the reasoning holds and the sources exist
Logging	which question, which model, which sources and who approved it

How RAG (retrieving knowledge) and MCP (taking actions) complement each other is covered in Beyond the buzzwords: MCP and RAG.

How do you measure reliability? With benchmarks: fixed tests on which models are scored. They tell you a lot, but not everything. For the Vals AI Legal Research Bench, practising lawyers wrote the questions and the grading criteria. A language model grades the answers against them. In September 2026 three models shared first place there, with 55.29% of answers passing every grading criterion. The standard error is about 3.5 percentage points, so the top eight models are statistically indistinguishable. The hardest part is reconciling conflicting sources: pooled across all models tested, only about a fifth of the answers there were fully correct. And that test covers US law.

An independent, peer-reviewed benchmark for legal research in Dutch, based on Dutch law, does not exist at this moment. That is why we are building our own. There are no results from it yet. On PrudentBench we track how models score on the existing tests, and which models pass our compliance gate.

Where should your organisation start with generative AI?

Start with the work, not the tool. Three questions get you further than a trial subscription:

Which work takes a lot of time without requiring real expertise? Searching, summarising, retyping. That is where the gain is, not in the judgement itself.
Which source must the answer rest on? No source, no use case.
Who checks, and how do you record it? Without an answer, the use case is not finished.

Also ask any supplier:

Can I click through from every answer to its source?
Where is our data stored, is it used for training, and how long is it kept?
How is human review set up, and what is logged?
What happens when the underlying model is replaced?
What does it cost per user per month, and what is included?

Checklist of five questions for a supplier: clicking through to the source, where data is stored and for how long, human review, what happens with a different model, and the cost per user per month.

Make sure your people know what a model can and cannot do. The AI Act asks providers and deployers of AI systems to take measures that support the AI literacy of their staff (Article 4). Read more in AI literacy under the EU AI Act.

How we work at Prudai

Three choices in our AI colleagues follow directly from how generative AI works.

Source first, answer second. LEO, VERA, ZIA, BEVER and IRMA work on one shared source registry with more than a hundred validated sources. Part of it is queried live at the source itself. Another part is indexed as a knowledge base. The public sources are listed on our knowledge sources page.
A person decides what gets recorded. In IRMA, AI proposals for risks, controls, process steps and draft policy only enter the register after a staff member has approved them. Every approved change is in the audit log, and AI proposals are off by default. ZIA can write back exactly one thing to the Nedap ONS care record system, a daily report, and always asks the care professional to confirm first. That connection is also off by default.
Not every strong model gets in. Before a model may run in production with us, it has to pass the compliance gate: contractual EU processing, an acceptable chain of processors and no vendor-mandated data retention that cannot be switched off, such as the 30 days Anthropic requires for Claude Fable. Some models that score highly do not pass.

We also published rechtspraak-mcp as open source under the MIT licence. It lets an AI assistant fetch rulings directly from the Dutch judiciary's Open Data instead of inventing them. You can search the rulings yourself at prudai.com/jurisprudentie.

For specialists

Context window. The number of tokens a model can take in at once. What falls outside it, the model does not see. Long files are processed in parts or retrieved on demand.
Temperature. The setting that controls how much randomness goes into token choice. Lower is more predictable, higher is more varied.
Open weights are not open source. With open weights the parameters can be downloaded, but the training data and recipe usually cannot. You can, however, run such a model on your own infrastructure.
RAG or live connection. RAG works with an indexed copy, which can go stale. A live connection queries the source at that moment. For legislation and case law, being current is often decisive.
Article 50 AI Act. Providers must mark synthetic text, images, audio and video in a machine-readable way as artificially generated. The obligation applies from 2 August 2026. Systems already on the market before then have until 2 December 2026 (Regulation (EU) 2026/1744). There are exceptions, for example for AI that only assists with standard editing. More in Article 50 of the AI Act.
High-risk AI. Under the same amending regulation, the requirements for high-risk systems in Annex III of the AI Act apply from 2 December 2027. For systems under Annex I the date is 2 August 2028.

Frequently asked questions

What is generative AI in one sentence?

Generative AI is artificial intelligence that creates new content, such as text, images, audio or code. It does so based on patterns learned from large amounts of existing data.

What does GenAI stand for?

GenAI is short for generative artificial intelligence. It means the same as generative AI.

Is ChatGPT generative AI?

Yes. ChatGPT is an application built on a generative language model. Generative AI is the underlying technology, and it also sits inside many search, writing and reporting tools.

Why does generative AI sometimes make things up?

The model is trained to produce plausible text, not to check facts. When it lacks the knowledge, it still gives the most likely answer, sometimes with a reference that does not exist. Connecting it to real sources and having a person review the output reduce that risk.

Is generative AI safe for confidential data?

That depends on how it is set up, not on the technology. Ask where the data is stored, whether it is used for training, how long it is kept and who can access it. Put the answers in the contract. Which routes exist and which agreements to make is covered in Is AI safe? And how do you keep control over your own data?.

Sources

Want to know how generative AI works when it is tied to real sources and a person has the final say? See Prudai's AI colleagues.

Updated on 2 October 2026

What is generative AI? Meaning, how it works and examples