Two Kinds of AI Walk Into a Bar
One of them looks at the bar and says, "That's definitely a bar. 98% confidence." The other one writes a short story about it.
That's the whole lesson, honestly. But stick around, because the details are fun.
For most of AI's history, the useful stuff was discriminative AI: systems that look at something and sort it. Is this email spam or not spam? Is this photo a cat or a dog? Is this transaction fraud or just you buying concert tickets at 2 a.m.? Discriminative AI draws lines between categories and files things on the correct side of the line.
Generative AI does something different: it creates new stuff that didn't exist before. New sentences, new images, new songs, new code. Instead of answering "which bucket does this belong in?", it answers "what would a plausible new thing look like?"
The Vacation Email Test
Here's a real-world way to feel the difference.
- Your spam filter reads an incoming email and makes one decision: inbox or junk folder. That's discriminative: it sorts an existing thing.
- Now you ask an AI assistant to write your out-of-office reply before you leave for vacation. It produces brand-new sentences that are polite, professional, maybe even a little charming. That's generative: it creates a new thing.
Same inbox. Two totally different jobs.
| Discriminative AI | Generative AI | |
|---|---|---|
| Core question | "Which category is this?" | "What new thing fits here?" |
| Output | A label or a score | Text, images, audio, code... |
| Example | Spam filter, face unlock | ChatGPT, image generators |
| Vibe | Bouncer checking IDs | Novelist on a deadline |
Your phone's face unlock is discriminative (is this YOUR face? yes/no). An app that generates a cartoon avatar of your face is generative. Both stare at you. Only one gets creative about it.
Why Everyone Suddenly Cares
Discriminative AI has quietly run the world for years: recommendations, fraud detection, autocorrect deciding you meant "ducking." But sorting things is invisible work. Nobody throws a party for a spam filter.
Generative AI got famous because its output is visible and human-shaped. When a machine writes a poem about your cat or drafts your cover letter, you can't not notice. The technology crossed a line from "useful behind the scenes" to "doing things we thought only people could do."
One system can wear both hats. A model can classify your question (discriminative-ish) and then generate an answer to it. The distinction is about the task, not team jerseys.
See the Difference in Action
Here's the same spammy email handled by both kinds of AI. Notice the shape of each output: one is a label, the other is new content.
INPUT EMAIL:
"CONGRATS you won a FREE yacht, click here!!!"
DISCRIMINATIVE AI (the sorter):
Output: SPAM (confidence: 99.7%)
That's it. One label. Job done.
GENERATIVE AI (the maker):
Output: "Free yacht in my inbox.
The sea of lies is boundless.
Delete, and sail on."
A brand-new spam haiku that never existed before.
Want to feel it for yourself? Open any AI chatbot and try both jobs back to back: first ask it to classify something ("Is this message spam?"), then ask it to create something ("Write a polite out-of-office reply for my beach vacation"). Same tool, two completely different kinds of output: a verdict versus a creation.
The One-Sentence Takeaway
Discriminative AI answers questions about things that exist. Generative AI makes things that don't. Yet.
Keep that sentence in your pocket. The rest of this course is basically an increasingly detailed answer to "okay, but HOW does it make new things?"
Checkpoint
Discriminative AI sorts existing stuff into categories; generative AI creates brand-new stuff.
Quick Quiz
๐ Quiz
Question 1 of 3Your bank flags a suspicious credit card charge. Which kind of AI is doing that?