Is every large language model generative?

Almost always, yes — an LLM like GPT or Claude is generative by design, since its entire job is producing new text one token at a time. But the same transformer building blocks can also be trained for discriminative tasks, like classifying whether a review is positive or negative. The generative/discriminative distinction is about what the model was trained to do, not just what architecture it uses.

Is discriminative AI outdated now that generative AI exists?

No — it's still the better and cheaper tool for most classification problems. Fraud detection, spam filtering, medical image triage, and credit risk scoring are almost all still discriminative models, because they need a fast, well-calibrated yes/no or category answer, not a creatively generated response. Generative AI didn't replace discriminative AI; it opened up a category of tasks discriminative models were never designed to do.

Can a generative model also do classification?

Yes, indirectly. You can ask an LLM 'is this email spam?' and get an answer, because generating the word 'yes' or 'no' is still just text generation to the model. But a dedicated discriminative classifier is typically faster, cheaper to run, and more consistent for narrow, high-volume classification tasks than using a full generative model for the same job.

What Is Generative AI? The Real Difference From Discriminative AI

If you had described “AI” to most people in 2015, they’d have pictured something that sorts or judges things: a spam filter, a fraud alert, a face-tagging feature on a photo app. Ask someone to describe AI today, and they’ll picture something that makes things: a chatbot writing an email, an image generator producing a picture from a text prompt. That shift has a name — it’s the move from discriminative AI to generative AI — and understanding it is the key to understanding why AI suddenly feels so different than it did a decade ago.

Discriminative AI: learning where the boundary is

A discriminative model’s entire job is to answer a “which category” question. Given an input, it learns to draw a boundary that separates one class from another, and then tells you which side of the boundary a new example falls on.

Think of a spam filter again: given an email, is it spam or not spam? Or a bank’s fraud detection system: given a transaction, is it fraudulent or legitimate? Or the classic example, is this a photo of a cat or a dog? In every case, the model isn’t trying to understand everything about the email, transaction, or photo — it’s specifically learning what separates the categories from each other. It’s not learning what “cat-ness” fundamentally is; it’s learning what specific features reliably tell cats and dogs apart.

This is the model of AI that dominated the 1990s through the mid-2010s: recommendation systems, ad-click prediction, credit scoring, search ranking. These systems are genuinely powerful and, importantly, they’re still what run most of the AI infrastructure quietly working behind the scenes today. They just don’t make headlines, because their output is a label or a number, not a paragraph or a picture.

Generative AI: learning the whole shape of the data

A generative model has a fundamentally more ambitious goal: instead of just learning the boundary between categories, it learns the underlying statistical structure of the data itself — well enough that it can produce brand-new examples that plausibly could have come from that same data, but didn’t.

Here’s the honest, technical version of what that means. When a language model is trained, it isn’t shown “this is a good sentence” and “this is a bad sentence” the way a spam filter is shown labeled examples. It’s trained to model the probability of what word is likely to come next, given everything before it — across a staggering volume of real text. Do that well enough, at large enough scale, and the model has effectively learned the statistical shape of human language: which words, phrases, and ideas tend to follow which others, in which contexts. Generating a new sentence is just sampling from that learned probability distribution, one token at a time.

Image generation models (like diffusion models, the technique behind most modern AI image generators) work on a related principle but reversed: they’re trained by taking real images and progressively adding random noise until the image is pure static, then learning to reverse that process — predicting, step by step, how to remove noise and recover something image-like. Once trained, you can start from pure random noise and run that denoising process guided by a text prompt, and the model will “denoise its way” toward an image that plausibly matches the description, even though that exact image never existed in the training data.

Neither of these processes involves the model classifying anything against a fixed set of labels. That’s the real, technically accurate difference: a discriminative model draws a boundary in existing categories; a generative model learns a distribution well enough to sample new points from it.

Why this shift is what unlocked the current AI boom

The practical consequence of this distinction is enormous, and it explains the timeline of AI products fairly precisely. The discriminative era produced AI you interacted with indirectly — you didn’t “talk to” your bank’s fraud model or your email’s spam filter; they ran silently and just changed what you saw. There was no reason for a consumer product built around “here’s a category label” to feel exciting to a general audience.

Generative AI is different because its output is the kind of thing people actually want directly: a written paragraph, a piece of code, a picture, a song. That’s why ChatGPT, Midjourney, and their successors could become mass-market consumer products in a way that fraud-detection models never could — not because generative techniques are “smarter” in some abstract sense, but because generating usable content is inherently more visible, shareable, and immediately useful to an individual person than sorting things into bins.

Discriminative AI didn’t go away — it’s still doing most of the invisible work

It’s worth being precise here instead of overselling the generative side: discriminative models remain the right tool for the vast majority of narrow, high-volume classification tasks, and they’re usually cheaper, faster, and more predictable than reaching for a generative model to do the same job. A bank isn’t going to replace its fraud-detection pipeline with an LLM asking “does this look fraudulent?” for every transaction — a purpose-built discriminative classifier will be faster, cheaper at scale, and more consistent. Generative AI didn’t make discriminative AI obsolete; it opened up an entirely different category of task — creation instead of classification — that the discriminative era was never built to handle.

Where this leaves the AI landscape

Put the last two articles in this series together and the picture becomes clear: machine learning is the umbrella, deep learning is the technique that made large-scale pattern learning practical, and within that, models split into discriminative ones that sort and generative ones that create. Today’s most talked-about AI products — chatbots, image generators, coding assistants — sit at the intersection of all three: deep learning models, trained generatively, built on transformer or diffusion architectures. But increasingly, the most capable of these systems aren’t just generating isolated pieces of text or images anymore — they’re being wired up to take actions, call tools, and pursue multi-step goals on their own. That’s the leap from a model that generates to a system that acts, and it’s exactly what the next article in this series digs into: how AI agents actually work.