What is generative AI? A plain-language guide

The short answer

Generative AI is software that creates new content — text, images, audio, video or computer code — in response to a request written in everyday language. It learned to do this by studying enormous collections of existing content and picking up the patterns in it. When you ask it for something, it produces a new piece that fits those patterns.

• **It generates, it doesn't look things up.** Unless it is connected to search or your documents, it writes from patterns, not from a database of facts.
• **It is very useful for drafts.** Summaries, emails, plans, explanations, translations and first versions of almost anything.
• **It can be confidently wrong.** Checking what matters is part of using it — always.

How it works, in four steps

1. **It reads a huge amount.** A [large language model](/glossary/large-language-model) is first trained on a vast collection of text — books, websites, code, articles. This [pre-training](/glossary/pre-training) takes months on thousands of specialised chips.
2. **It learns to predict the next piece.** During training it plays one game billions of times: guess the next small chunk of text (a [token](/glossary/token)), check, adjust. Getting very good at that game turns out to require learning grammar, facts, styles and some reasoning.
3. **People shape its behaviour.** Raw models are then trained further with human feedback ([RLHF](/glossary/rlhf)) and written rules, so they follow instructions, decline harmful requests and answer in a helpful tone.
4. **Your prompt starts the generation.** When you type a [prompt](/glossary/prompt), the model produces its answer one token at a time, each choice based on everything before it. Image generators work differently — usually with [diffusion models](/glossary/diffusion-model) that turn noise into a picture — but the idea of learning patterns and generating something new is the same.

A useful mental model: generative AI is like a well-read assistant with a huge memory for how things are usually said, no memory of where it read them, and a strong urge to give you an answer even when it isn't sure.

What it can make

Generative AI vs the AI you already use

AI was part of daily life long before chatbots. Spam filters, fraud alerts, map traffic estimates and photo search use [machine learning](/glossary/machine-learning) to classify or predict. Generative AI uses similar foundations but produces new content instead of a label or a number.

The main tools

Most people meet generative AI through a general [AI assistant](/glossary/ai-assistant). All of these have free plans and paid plans with stronger models and higher limits. Features change often; this list was reviewed in October 2026.

• **ChatGPT** (OpenAI, United States) — the best-known assistant; text, images, voice and web search.
• **Claude** (Anthropic, United States) — strong at writing, long documents and coding.
• **Gemini** (Google, United States) — built into Google's apps, Android and Workspace.
• **Copilot** (Microsoft, United States) — built into Windows, Edge and Microsoft 365.
• **Le Chat** (Mistral AI, France) — a European option that works well in French.
• **Open-weight models** such as Llama, Mistral, Gemma and Qwen can run on your own computer — more private, more technical. See [open-weight model](/glossary/open-weight-model).

Which one is "best" matters less than practice. Pick one, use it for a week on real tasks, and learn its habits. Our free [Labs path on AI assistants](/labs/assistants) compares them side by side.

What it gets wrong

• **Made-up facts.** [Hallucinations](/glossary/hallucination) — invented quotes, studies, laws and court cases — are the most common serious failure. Ask for sources and open them.
• **Out-of-date knowledge.** Without web search, a model knows nothing after its [knowledge cutoff](/glossary/knowledge-cutoff): new prices, rules or events.
• **Bias.** It can repeat stereotypes in its [training data](/glossary/training-data), and it often serves French and other languages less well than English.
• **Telling you what you want to hear.** Models trained to please can agree with a wrong premise. Ask it to argue the other side.
• **Maths and exact details.** Long calculations, dates and counts can slip unless the tool uses a calculator or code.
• **Privacy.** What you type may be stored and, depending on settings, used to train future models. See our guide to [using AI assistants safely](/guides/use-ai-assistants-safely).

What it means for people

We judge every technology story by what it changes for people. Generative AI has real benefits and real costs, and they don't fall on everyone equally.

• **Work.** It changes tasks in office, creative and technical jobs first. Read [AI and jobs in Canada](/guides/ai-and-jobs-in-canada).
• **School.** It can be a patient tutor or a shortcut that skips the learning. Read [AI for students](/guides/ai-for-students).
• **Access.** Read-aloud, captions, translation and plain-language rewrites help people with disabilities, newcomers and anyone facing dense paperwork.
• **Trust and democracy.** Cheap, convincing fakes make [misinformation](/glossary/misinformation) and [deepfakes](/glossary/deepfake) easier to spread.
• **Creators.** Writers, artists and journalists are asking courts whether training on their work without permission is legal. See [AI and copyright](/glossary/ai-and-copyright).
• **The environment.** Training and running large models uses a lot of electricity and water in [data centres](/glossary/data-centre).

Generative AI in Canada

Canada helped invent the deep-learning methods behind today's generative AI: researchers in Toronto, Montréal and Edmonton — home to the Vector Institute, Mila and Amii — did foundational work, and Canada launched the world's first national AI strategy in 2017. The federal government now has a minister responsible for artificial intelligence; our [AI Ministry tracker](/ministry) follows every announcement.

Use is growing fast. Statistics Canada's business surveys found the share of businesses using AI to produce goods or deliver services roughly doubled between 2024 and 2025, to about one in eight — still a minority, concentrated in information, professional and finance sectors. There is no AI-specific federal law yet (the proposed [AIDA](/glossary/aida) died in 2025), but privacy, human-rights, consumer and employment laws already apply, and Québec's [Law 25](/glossary/law-25) sets strict privacy rules.

How to start in 15 minutes

1. Open one assistant (ChatGPT, Claude, Gemini, Copilot or Le Chat). Turn off training on your chats if you prefer — see [the safety guide](/guides/use-ai-assistants-safely#settings).
2. Give it a real task with context, a goal and a format: "I'm a parent of two. Plan five weeknight dinners under 30 minutes, nut-free, with a grocery list."
3. Push back: "Make it cheaper", "Explain step 3", "What might be wrong here?"
4. Check one fact it gave you against a trusted source. Make that a habit.
5. Keep going with our free guide [Start using AI in 30 minutes](/learn/start-in-30-minutes) or the [Start here Labs path](/labs/start-here).