Predicting an answer for new data
Predictive AI learns regularities from existing data and produces a value, a classification, or a likelihood for data it has not seen.
It is not only about the future. Whether an email is spam, what a photograph shows, which product a customer is likely to prefer — producing an answer for data that does not have one yet is prediction too.
Everyday examples
- Forecasting weather, sales, demand, or prices
- Flagging spam, fraudulent transactions, or likely equipment failure
- Classifying what is in a photograph or what kind of text something is
- Recommending products, music, or video likely to suit someone
- Estimating the probability that a customer will churn
Most of these train a model on historical data where the answer is known. The inputs are called features and the target is called the label.
How it differs from generative AI
Predictive AI is good at producing a result from a set of options or a numeric range. Generative AI is good at producing new content — text, images, audio, code.
The boundary is not clean. Generative models also write by predicting the next token probabilistically. And the two combine: generative AI can explain a predictive model’s output, or draft a proposal from information a predictive model selected.
A prediction is not a certainty
What predictive AI gives you is a likelihood computed from training data. Unprecedented events, changing conditions, and skewed data all degrade accuracy. A figure like 90% also needs the conditions it was measured under.
For consequential uses such as hiring, lending, and healthcare, watch for bias and blind spots and keep a path for human judgment. Which error matters more — false positives or missed cases — also depends on the application.