One forecasts a number or sorts a case into a category it has already seen before. The other produces something new that did not exist as an input. Knowing which one a tool actually is changes what you should expect from it, and what kind of mistake it can make.
Key takeaways
The Canadian Centre for Cyber Security states the distinction directly: 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. “Traditional” here is the predictive style of AI — scoring, sorting, forecasting against categories the system already has. Generative AI is defined by contrast, as the kind that creates something that was not one of a fixed set of pre-existing options.
Worth noting: the word “traditional” in that sentence is doing real work. It signals that predictive, pattern-recognizing AI is the older, more established style of the technology, not a lesser one — credit scoring, fraud detection and demand forecasting have run on predictive models for years, well before the current wave of generative tools, and remain, in many businesses, the more thoroughly tested of the two approaches.
A predictive system is trained on historical examples that each carry a known outcome — a loan that did or did not default, a support ticket that was or was not urgent, a month's demand that came in at a certain level. From that, it estimates the odds a new case belongs to one of those known categories, or produces a number within a learned range. Its output is bounded by what the label set could ever contain: a fraud-scoring model can say a transaction is more or less risky, but it cannot say something outside the categories it was built to sort into.
That boundedness is also what makes a predictive system's errors countable in a way few other kinds of software failure are. Because every prediction can eventually be checked against what actually happened — the loan defaulted or it did not — a business can build a real error rate for a predictive tool: how often it flagged a good case as bad, and how often it missed a bad one. That measurability is a genuine strength of the predictive style, not a footnote to it.
A generative system is not bounded by a fixed label set at all. That is precisely why it can answer a question nobody explicitly trained it to answer — and precisely why it can also produce a fluent, structurally plausible wrong answer in a way a bounded classifier cannot. A spam filter can only ever say “spam” or “not spam;” it cannot invent a third, wrong category. A generative model has no such ceiling, for better and for worse.
This is also why the error-rate math from the predictive side does not transfer cleanly. There is no fixed set of possible generated answers to check a given output against, so “how often is it wrong” is a much harder question to define, let alone measure, for a generative system than for a classifier with a known set of labels.
Ask what the possible outputs are before the system runs. If you can list every possible answer in advance — approve or decline, urgent or not urgent, a number between zero and some maximum — you are almost certainly looking at a predictive system, however it is marketed. If the possible outputs are effectively unlimited — any sentence, any image, any block of code — you are looking at a generative one. This single question cuts through most marketing language faster than reading a spec sheet, because a vendor can call either style “AI-powered” or even “smart” without being specific about which mechanism is underneath.
A second, related question worth asking is what the system is trained to optimize for. A predictive model is typically trained to get as close as possible to a known correct answer on historical examples — its whole training process is organized around minimizing the gap between its prediction and what actually happened. A generative model is typically trained to produce output that continues a pattern plausibly, which is a genuinely different objective, and one reason the two styles fail in different ways even when both are described, loosely, as making a prediction.
Statistics Canada's most recent survey of AI-using businesses gives a real, dated picture of the split. Among Canadian businesses that had used AI in the past 12 months, 13.7% reported using AI for decision-making systems and 17.9% for recommendation systems in the second quarter of 2026 — both closer to the predictive style — against 24.8% reporting use of large language models. These are usage figures, not performance claims: they describe what Canadian businesses have actually adopted, not which approach works better.
A system that predicts which lead is most likely to convert, and then drafts the outreach message to the leads it ranked highest, is stitching a predictive mechanism to a generative one inside a single product. Each half fails differently: the predictive half can be miscalibrated against a category it has too little data on, while the generative half can produce a fluent message that gets a fact wrong. When evaluating a tool that does both, it is worth asking which specific step is doing which job, since can AI cite its sources reliably and machine learning vs AI vs deep learning address in more detail what to check on the generative half once you have identified it.
The practical consequence for a buyer is that a single vendor demo covering both halves can hide which one to be more cautious about. Ask separately: how is the predictive ranking validated, and against what outcome data, and separately again, how is the generated message checked before it reaches a customer. A tool that answers the first question with real numbers and waves off the second with “it's just drafting” has told you exactly where its own testing stopped.
See also generative AI vs agentic AI and machine learning vs AI vs deep learning.
Predictive. It classifies each message into one of a fixed set of categories based on patterns learned from labelled examples — it does not produce new content, and it could not invent a third category even if one would fit better.
It can be both at once: the underlying forecast may come from a predictive model, and a generative layer turns the number into a written sentence explaining it. Two different mechanisms can sit behind one piece of output, and each should be checked on its own terms.
No. In a predictive system, a false positive is a defined, countable error against a known label — you can measure a false-positive rate. A generative system's wrong answer often has no equivalent labelled ground truth to test against, which is a different kind of failure to manage; see can AI cite its sources reliably for that specific failure mode.
Deciding which of the two a specific build needs, and what it should cost, is a scoping question. For that, see five questions to ask before approving a custom AI budget.
