Canada’s Cyber Centre supplies the cleanest government definition available: “Generative AI is a type of AI that generates new content by modelling features from large datasets that were fed into the model.” The line that actually separates it from older software is the sentence right after: “While traditional AI systems can recognize patterns or classify existing content, generative AI can create unique content in many forms, including text, image, audio or software code.” A spam filter that sorts mail into folders is not generative. A tool that drafts the reply is.
Ottawa’s own Voluntary Code of Conduct on advanced generative AI systems reaches for examples rather than a formal definition, and its examples are worth reading because they are what the government itself was reacting to: “Advanced AI systems capable of generating content — such as ChatGPT, DALL·E 2, and Midjourney — have captured the world’s attention.” It goes on to describe what these systems actually do day to day: “These advanced systems may be used to perform many different kinds of tasks — such as writing emails, answering complex questions, generating realistic images or videos, or writing software code.” Signing the Code is voluntary. It binds only the organizations that have signed it — it is not a general Canadian law that tells anyone how a generative AI system has to be built or labelled.
The Cyber Centre’s guidance treats the most familiar version of the category as one slice of it, not the whole of it: “A subset of generative AI that has seen significant improvement in recent years is large language models (LLMs).” A large language model is the text-focused case most people mean when they say “generative AI”; image, audio and code generators sit in the same category built on the same idea, usually starting from a foundation model trained the same general way.
A mortgage brokerage’s after-hours chat widget makes the line concrete. If the widget only matches a client’s typed question against a fixed list of pre-written answers and returns the closest match word for word, nothing about it is generative — it is a lookup table with a search box in front of it. The same widget becomes generative the moment it composes its own sentence for a question nobody wrote an answer to in advance, drawing on patterns it picked up during training rather than a line someone typed into a spreadsheet. Almost everything built on top of a chat model today is the second kind, which is exactly why the Cyber Centre’s caution about validating output before acting on it applies to it and not to the lookup table.
See also large language model, foundation model and AI hallucination for what generative systems build on and where they go wrong.
This is one term in a plain-English glossary on how AI actually works and where it fits in a Canadian business.
