Definition
One AI tells you a photo shows a cat. Another AI draws a cat that never existed. What is the difference?
Predictive AI studies data and gives a judgement about it: a label, a number or a ranking. Generative AI learns the patterns in data and uses them to produce new content, such as text, images, audio or video. Both learn from examples; they differ in what comes out.
In simple words
A cricket umpire and a cricket commentator both know the game. The umpire gives a decision: out or not out. The commentator creates sentences nobody has said before. Predictive AI is the umpire. Generative AI is the commentator.
Why it matters
The two kinds fail in different ways, so they need different checks. A predictive model can put something in the wrong group. A generative model can produce something that looks right and is false. Knowing which kind you are using tells you what to watch for.
Key terms
- Predictive AI
- AI that gives a judgement about data: a label, a number or a ranking.
- Generative AI
- AI that produces new content such as text, images, audio or video.
- Label
- A category that a predictive model assigns, such as "spam".
- Content
- Material made for people to read, see or hear.
- Prompt
- The instruction or question given to a generative AI tool.
- Plausible
- Looking or sounding right, whether or not it is true.
Explained in detail
The older, quieter kind
Predictive AI has been at work for decades, and it still does most of the AI work in the world. It decides which emails are spam, which payments look suspicious, which route is fastest, which patients need attention first and which products to show you. Each of these is a judgement about something that already exists. Predictive AI rarely makes headlines, because when it works you do not notice it.
How creating grows out of predicting
Generative AI looks like a different thing, but it is built from prediction. A text generator does one small job: given the words so far, predict the next piece of text. It then adds that piece and predicts again, and again, hundreds of times, until a paragraph exists. An image generator starts with a field of random dots and repeatedly predicts how to remove a little of the randomness so that the picture moves closer to the description. In both cases, a long series of small predictions adds up to something new.
Why generative AI arrived when it did
Generating a convincing paragraph or picture is much harder than choosing a label, because there are endless possible outputs and most of them are nonsense. It became practical only around 2020, when three things were in place: a network design that could handle long passages of text, training collections containing a large part of the public internet, and enough specialised chips to train models with many billions of parameters. Generative models are far larger and costlier to run than most predictive ones.
Judging them is different
A predictive model can be scored. There is a right answer for each test case, so you count how often the model matched it. A generative model has no single right answer. Ten different poems about the monsoon may all be good. People have to judge whether the output is useful, accurate, safe and well made, and they often disagree. This is one reason generative AI is harder to trust: it is harder to measure.
| If you need to… | Use | Because |
|---|---|---|
| Sort 10,000 complaints by urgency | Predictive AI | A clear label for each, checked against known answers |
| Estimate next month's electricity demand | Predictive AI | A number, measured against what really happens |
| Draft a polite reply to each complaint | Generative AI | New text is needed, and a person can check it |
| Make a poster for the school fair | Generative AI | A new image, with many acceptable results |
They work together
Real systems mix the two. A hospital may use a predictive model to flag a scan and a generative model to draft the report that the doctor then corrects. A chatbot's reply may pass through a predictive filter that blocks harmful content before you see it. Knowing which part is which tells you what to check: for the predictive part, whether the label is right; for the generative part, whether the content is true.
How it works
| Predictive AI | Generative AI | |
|---|---|---|
| Question it answers | "What is this?" or "What will happen?" | "Make me something new" |
| Output | A label, a number, a yes or no, a ranking | A paragraph, a picture, a song, a video |
| Examples | Spam filter, fraud alert, weather forecast, face unlock | Chat assistants, image makers, music makers |
| Right answers | Usually one correct answer | Many acceptable answers |
| Typical mistake | Wrong label or wrong number | Fluent content that is untrue |
The two are closer than they look. A chat assistant writes by predicting the next small piece of text again and again. So generation is built from many tiny predictions. The difference is what you receive at the end: a judgement about something that exists, or a new thing.
Try it yourself
The same photos, two different models
Both models learned from thousands of photos of flowers.
Predicts or creates?
Look at what each tool gives back.
Draws a logo from a description
Decides whether a payment looks like fraud
Composes background music for a video
Writes a poem about the monsoon
Estimates tomorrow's temperature
Marks an email as spam
Common mistakes
- Thinking generative AI is a completely different technology. It is built from predictions.
- Thinking generated content is checked for truth. It is produced to look plausible.
- Thinking predictive AI is old and no longer used. It runs most of the AI around you.
Check yourself
Summary
- Predictive AI judges: it gives a label, number or ranking.
- Generative AI creates: it produces new text, images, audio or video.
- Both learn patterns from data, and each has its own kind of mistake to check for.
Finished reading?
Sources: Elements of AI (University of Helsinki) · AI4K12: Five Big Ideas in AI