This is not a tour of the mathematics. It is a sequence — ten stops, in the order they actually happen to a request you type — built so that by the end you can place any AI product you encounter into the picture, rather than treating each new tool as a fresh mystery.
Key takeaways
STEP 01 OF 11
Every AI system has a training phase — when it learned patterns from a large body of data — and an inference phase, or every individual time it is actually used afterward. These happen at completely different times and cost structures: training is expensive, done rarely, and finished long before your question exists; inference is what happens in the seconds after you type something in.
This split explains a lot on its own. The system cannot learn anything new from your conversation the way a person would — by default, what it knows was fixed when training ended, and your specific question is not added to that knowledge unless a tool is deliberately built to do so.
STEP 02 OF 11
Canada’s Cyber Centre defines generative AI as a system that “generates new content by modelling features from large datasets that were fed into the model” — from ITSAP.00.041. “Modelling features” means the system absorbed statistical regularities in how words, images, or sounds tend to go together — not the underlying facts those patterns describe.
This is why a system can produce fluent, well-structured text about something it does not actually understand in any deeper sense: it learned the shape of correct-sounding answers, which usually but not always coincides with an actually correct one.
STEP 03 OF 11
Once you send a request, the system builds its answer one small unit at a time — a word or word-fragment, called a token. Google DeepMind’s own technical description of this step, from its SynthID documentation: “Large language models generate text one word (token) at a time. Each word is assigned a probability score, based on how likely it is to be generated next.” For a sentence like “My favourite tropical fruits are mango and…”, the word “bananas” scores higher than the word “airplanes.”
Note what this is describing: a system predicting the next plausible word given everything before it, not consulting a stored fact and reporting it back. That single distinction explains most of what is unusual about how these tools succeed and fail.
STEP 04 OF 11
Because each word is chosen for being statistically likely rather than verified as correct, a wrong answer and a right one can be generated through exactly the same process, in exactly the same tone. Nothing in the token-by-token mechanism itself flags “I am not sure about this part.”
See how to sanity-check an answer from AI for what to actually do about this — the short version is that fluency and accuracy are unrelated properties of the same output.
STEP 05 OF 11
A plain chat response draws only on what the model learned during training, plus whatever you typed in this conversation. Many tools now add a further step: the system can call an external tool — a search, a database lookup, a calendar — and use the result before finishing its answer. See what an integration does in plain terms for that mechanism specifically.
This distinction matters for how much you should trust an answer about anything recent or specific to your own records: a model with no connection to today’s data or your own systems is, by construction, answering from what it learned before training ended.
STEP 06 OF 11
A further step up is a system that does not just answer, but takes an action in another system on its own — sends the email, updates the record, books the appointment. See what an AI agent actually does for the specific dividing line: an agent can call a defined tool and act on what it observes, where a chat response can only describe what it would do.
Recognising which category a tool falls into — answers only, or answers and acts — changes how much checking that tool needs before you rely on it, which is the subject of a separate guide entirely.
STEP 07 OF 11
The Cyber Centre’s definition again is the cleanest split: traditional AI “can recognize patterns or classify existing content”; generative AI “can create unique content in many forms, including text, image, audio or software code” (ITSAP.00.041). A spam filter and a chatbot that drafts a reply are both “AI”, doing structurally different jobs.
Knowing which category a tool sits in tells you what kind of mistake to expect from it: a classifier can mis-sort something; a generator can invent something that was never there at all. These are different failure modes and deserve different scrutiny.
STEP 08 OF 11
The step-by-step mechanism above — predicting the next token from patterns in training data — is not new. What changed, in a short span, was the scale of training data and computing power applied to it, which produced a jump in how fluent and broadly capable the output became. That timing is why a technology built on a decades-old idea felt, to most people, like it appeared overnight.
It also explains why the tools keep changing quickly: the underlying mechanism is the same across versions, but each new version is trained on more data with more computing power behind it, which is a difference of scale, not of a fundamentally new idea.
STEP 09 OF 11
As of this guide, Canada has no AI-specific statute in force. Bill C-27, which would have enacted the Artificial Intelligence and Data Act, is recorded on LEGISinfo against the 44th Parliament’s 1st session (22 November 2021 to 6 January 2025) — a session the page itself marks as prior — with its last recorded status “At consideration in committee in the House of Commons.” What exists instead is a voluntary code industry signatories can opt into.
The full detail of what is and is not regulated belongs in how to track Canadian AI guidance — the point here is only to place the legal landscape correctly on the same mental timeline as the technology, so “is this legal” and “is this regulated” are not assumed to be the same question.
STEP 10 OF 11
Run any new AI product through the questions this sequence built: is it in a training phase or being used live right now? Is it working only from what it learned, or reaching an external source? Does it only answer, or can it also act? Is it classifying existing content or generating new content? Answering these four questions for a specific tool tells you more about what to expect from it than any marketing description will.
That is the actual point of working through the mechanism once: not to become able to explain how AI works from first principles, but to have a short, repeatable set of questions that place a new tool correctly the first time you encounter it.
STEP 11 OF 11
Newer versions of these systems train on more data with more computing power, per Step 8, but the underlying mechanism — predicting the next plausible token from learned patterns — has not changed. That means the practical habits this sequence points toward, especially checking an answer before relying on it, do not expire as the tools get better. A more capable system is a more convincing one, which is a reason for more scrutiny of confident-sounding answers, not less.
This is worth stating plainly because the opposite assumption is common: that a newer, more fluent tool needs less checking than an older one. Fluency has improved considerably; the underlying reason a wrong answer can sound exactly like a right one has not gone away.
Assuming the system checked a fact because it stated one confidently. The token-by-token mechanism in Step 3 produces fluent text regardless of whether the underlying claim is correct — confidence of tone carries no information about accuracy.
Assuming a chat tool knows about anything after its training ended. Unless a tool is explicitly built to reach outside itself, as in Step 5, it is answering purely from what it learned before training stopped — which can be months or longer before you are using it.
Treating “AI” as one thing. A classifier and a generator, or a plain chat response and an agent that acts on other systems, fail in different ways and deserve different levels of scrutiny — see Steps 6 and 7.
Assuming Canada regulates this the way it feels like it should. As Step 9 covers, there is no AI-specific statute in force in Canada today. Do not assert otherwise, and do not assume the absence of one means nothing applies — existing privacy and consumer-protection law still does.
Assuming a newer, more fluent tool needs less checking. Fluency has improved a great deal; the mechanism that lets a wrong answer sound exactly like a right one, covered in Step 4, has not gone away in any version of these systems.
These four answers, taken together, predict most of what a specific tool will be good at and where it is likely to go wrong — before you have used it even once.
Generative AI is built using machine-learning techniques, but not every machine-learning system is generative. A spam filter is machine learning that classifies; a chatbot is machine learning that generates new content. The mechanism in this guide describes the generative case specifically.
Only if the specific product is built to store and reuse that history — it is not a property of the underlying mechanism itself. By default, each fresh conversation starts from the same trained state, with no memory of a different one unless the product explicitly adds that feature.
They can be trained on different data, tuned differently, or reach different external sources, per Step 5. See why AI answers the same question differently for the fuller picture.
No — the point of working through this sequence once is to build a working mental model, not technical fluency. Knowing that answers are generated word by word from learned patterns, per Step 3, is enough to explain most of what is unusual about how these tools succeed and fail.
The next question is usually whether it is even worth doing, in what order.
