B4 · Domains of AI: Vision, Language, Speech and Robotics

Predictive AI and Generative AI

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.

8 min Beginner Jump to the activity

What you need: Nothing. This page starts from the beginning, with no programming and no formulas.

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…UseBecause
Sort 10,000 complaints by urgencyPredictive AIA clear label for each, checked against known answers
Estimate next month's electricity demandPredictive AIA number, measured against what really happens
Draft a polite reply to each complaintGenerative AINew text is needed, and a person can check it
Make a poster for the school fairGenerative AIA 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 AIGenerative AI
Question it answers"What is this?" or "What will happen?""Make me something new"
OutputA label, a number, a yes or no, a rankingA paragraph, a picture, a song, a video
ExamplesSpam filter, fraud alert, weather forecast, face unlockChat assistants, image makers, music makers
Right answersUsually one correct answerMany acceptable answers
Typical mistakeWrong label or wrong numberFluent 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

Question 1

1. Which is generative AI?

Question 2

2. What does predictive AI usually give back?

Question 3

3. What is the typical mistake of generative AI?

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