A few years back, if you told someone at work “we’re using AI for this,” it usually meant a fraud model, a churn predictor, or a recommendation engine quietly running in the background. Nobody was impressed by it in the way they are now. It just worked, and most people never saw it directly.
Then ChatGPT showed up, and suddenly “AI” meant something you could talk to. Something that wrote your emails, drew pictures, explained your code. Same three letters, very different feeling. So what actually changed?
Turns out, quite a lot. But the core idea isn’t complicated once you see it.
Two Very Different Jobs
Most AI systems before 2020 or so were built to do one job: look at something and make a decision about it.
- Is this email spam or not spam?
- Will this customer churn next month?
- What’s the probability this transaction is fraud?
- What will demand look like next quarter?
This is called predictive AI. It takes existing data, finds patterns in it, and outputs a label, a score, or a number. It’s not creating anything new; it’s classifying, scoring, or forecasting based on what it’s already seen.
Generative AI does something fundamentally different: it creates new content that didn’t exist before.
- Write a paragraph explaining OLTP vs OLAP
- Generate an image of a lighthouse at sunset
- Write a SQL query from a plain English description
- Draft an email based on three bullet points
Same underlying discipline (machine learning), completely different job. One predicts. The other produces.
A Simple Way to Tell Them Apart
Ask yourself: is the output a choice from a fixed set of options, or is it something new?
| Question | Predictive AI | Generative AI |
|---|---|---|
| “Is this transaction fraudulent?” | Yes / No | — |
| “What’s tomorrow’s stock price?” | A number | — |
| “Write a product description for this shoe” | — | New text |
| “Generate a chart summarizing this data” | — | New content |
If the model is picking from a small, predefined set of answers (categories, scores, numbers), that’s prediction. If it’s producing something open-ended (text, images, code, audio), that’s generation.
Seeing the Difference
I find this easier to show than to describe, so here’s the same idea laid out as a picture.
flowchart LR
A[Input: Email text] --> B[Predictive Model]
B --> C{Pick one label}
C --> D[Spam]
C --> E[Not Spam]
F[Input: 'Write a product description'] --> G[Generative Model]
G --> H[New text, created word by word]
The top row is the model you’ve probably worked with already: an email comes in, and it picks one answer from a small, fixed list. “Spam” or “Not Spam.” Nothing about that list changes; it existed before the model ever ran, the model just has to choose from it.
The bottom row is doing something genuinely different. A prompt comes in, and instead of choosing, the model builds, word by word, a piece of text that didn’t exist five seconds earlier.
Picking vs. building. That’s really the whole distinction.
Why This Distinction Matters
This isn’t just trivia. It changes how you evaluate and trust these systems.
A predictive model is either right or wrong in a fairly measurable way; you can check its accuracy against ground truth. A generative model doesn’t have a single “correct” answer. Ask it to write a poem about databases, and there are a thousand valid poems it could write. That makes evaluation harder, and it’s part of why generative AI systems need different guardrails than predictive ones, something we’ll get into in a later article.
Mental Model
Think of predictive AI as a very good judge: it looks at evidence and renders a verdict from a known set of outcomes. Think of generative AI as a very good writer: it takes a prompt and produces something original, drawing on patterns it learned, but never just picking from a list.
Same training process under the hood (both learn from data). Completely different jobs at the end.
Key Takeaways
- Predictive AI answers questions with a fixed set of possible outputs: classification, scoring, forecasting.
- Generative AI creates new content (text, images, code) that didn’t exist before the prompt.
- Both are built using machine learning, but they solve different problems and are evaluated differently.
- Understanding this distinction is the first step before diving into how generative models (like LLMs) actually work, which is exactly where we’re headed next.
Next up: How LLMs Actually Work, what’s really happening when a model generates text, one token at a time.