EXPLAINER

What is Generative AI?

Generative AI models create new content — text, images, audio, video and code — by learning the patterns in very large training sets.

Short answer: Generative AI produces new content that resembles its training data, instead of only classifying or predicting. Chat assistants, image generators and code assistants are all generative AI.

What is it?

Traditional ML is mostly discriminative: given an input, choose a label or number ("spam", "$312,000"). Generative AI models instead learn the structure of the data itself well enough to produce new samples — a paragraph, a picture, a melody, a function in Python.

Common families:

  • Language models generate text (and code) one piece at a time. See What is a Large Language Model (LLM)?.
  • Diffusion models generate images (and some video/audio) by starting from noise and repeatedly refining it towards something that matches a text description.
  • Multimodal models accept or produce more than one kind of content. See What is Multimodal AI?.

Many of these are built on a transformer architecture, although image generators commonly combine it with diffusion techniques.

Why does it matter?

For the first time, a general-purpose interface — plain language — can drive software across many tasks: drafting, summarising, translating, explaining code, extracting data from documents, brainstorming. That lowers the skill barrier for tasks that used to need specialised tools or people.

It also introduces new risks that older software did not have: fluent but wrong output, leaked sensitive data, copyright and attribution questions, and content that is hard to tell from human work. A useful mental model is fast, tireless, well-read assistant that sometimes makes things up — valuable, but needing review.

How does it work?

At a high level, training and use look like this:

  1. Pre-training on a very large collection of text, images or other data. For a language model the training task is simply to predict what comes next; to do that well across a huge variety of text, the model ends up capturing grammar, facts, styles and reasoning patterns.
  2. Adaptation — optional extra training so the model follows instructions and behaves helpfully and safely (see What is Fine-Tuning an AI Model?).
  3. Generation (inference) — given your prompt, the model produces output step by step. There is a controlled amount of randomness, so the same prompt can give different results.

The model does not look answers up in a database. It generates plausible output from patterns stored in its parameters, which is why it can be creative and also why it can be wrong. Techniques like What is RAG (Retrieval-Augmented Generation)? add real source material to the prompt to reduce that risk.

Example

A practical workflow for a support team:

  1. Paste a long customer email thread into an assistant.
  2. Ask for a three-bullet summary and a draft reply in a friendly tone.
  3. A human checks the facts (dates, refund amounts), edits, and sends.

The model saves reading and drafting time; the human remains accountable for accuracy. Removing step 3 is where most real-world problems begin.

When should I use it?

It works well for first drafts, summarising, rewriting, translation, classification of messy text, explaining unfamiliar code, brainstorming and transforming formats (notes to email, table to JSON).

Be cautious or add verification when exact facts, numbers, legal/medical/financial correctness or citations matter, when the output is published unreviewed, or when the input contains confidential data (see AI Privacy and Security: What to Know Before You Paste Data).

Common mistakes

  • Using it as a search engine or database. It generates; it does not retrieve, unless you add retrieval (What is RAG (Retrieval-Augmented Generation)?).
  • Believing confident tone means accuracy.
  • Pasting sensitive data into a service you have not vetted.
  • Expecting identical output every time. Generation is probabilistic.