In plain terms
Earlier AI mostly judged things: is this email spam, will this customer leave, what is in this photo. Generative AI makes things: it writes the reply, draws the picture, composes the music, produces the code. You describe what you want in ordinary language and receive a draft in seconds.
Why it matters
It moved AI from specialist teams to every desk, because using it requires neither training data of your own nor programming. Its economic effect falls on knowledge work: drafting, summarising, analysing, coding, searching. Its characteristic risk is new as well: the output is fluent and plausible whether or not it is correct. Organisations get value when they combine it with their own data, with checks on what it produces and with a clear idea of the work it is meant to change.
Example
A marketing team needs product descriptions for 3,000 items in four languages. Written by hand, that is about four months of work. A generative model produces first drafts from the product data overnight, and two editors spend three weeks reviewing and correcting them. The saving is real, and so is the need for review: roughly one description in fifteen contained a claim that the product data did not support.
Most often confused with
GenAI vs. Predictive AI
Predictive AI, also called discriminative or traditional AI, answers a narrow question with a score: fraud or not, how many units, which category. Generative AI produces open-ended content. They solve different problems and are often combined. For forecasting and scoring on structured data, predictive models remain more accurate, cheaper and easier to audit.
Origin: The technology developed over the 2010s; the term entered everyday use after ChatGPT was released in November 2022.
Under the hood
Main model families: large language models (transformers that generate text and code token by token), diffusion models (images, video, audio) and multimodal models that accept and produce several kinds of content. All are trained by self-supervised learning on very large datasets and then adapted through prompting, retrieval or fine-tuning. Output is sampled from a probability distribution, so the same request can give different results. Characteristic issues: hallucination, copyright and the provenance of training data and outputs, bias, prompt injection and the cost of inference. Typical enterprise patterns, in rising order of complexity: assistants, RAG over company knowledge, workflow automation and agents.