How do predictive AI and generative AI differ?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
Predictive AI estimates an outcome, a category or a numerical value from supplied information. Generative AI produces content such as text, images, audio or code. These terms describe what a system is used to do. They are not two completely separate mechanisms with no overlap.
A retailer might use predictive AI to estimate next week’s demand for umbrellas. A generative system might draft the product description or create an illustration for an advertisement. The first task asks for an estimate about something; the second asks for content. Both can depend on patterns learned from previous examples.
A prediction does not have to concern the future. Recognizing whether a photograph contains an umbrella is also a prediction in the technical sense: the system estimates the appropriate answer from the available information. Classification means predicting a category. Regression means predicting a number, such as expected sales.
The overlap becomes clearer with language models, systems trained to recognize and generate language patterns. Many generate text by repeatedly predicting a next token, which is a small piece of text such as a word fragment. Prediction is therefore part of the mechanism producing a generative result. Saying that a model “only predicts” does not, by itself, describe the range of tasks that this mechanism can support.
The same model can also be used in different ways. It might generate a paragraph in one application and assign customer messages to predefined categories in another. The product’s purpose and the form of the required answer matter as much as the underlying model family.
The standards for assessing results differ. A demand estimate can be compared with observed sales, while accounting for circumstances such as stock shortages. A fictional illustration may have many acceptable versions. A generated factual explanation needs checking for both correctness and whether it represents its sources accurately. Fluency is not a substitute for that evidence.
In a hospital discharge workflow, estimating a patient’s 30-day readmission risk from clinical measurements is a predictive task. Drafting a clear, plain-language discharge instruction sheet from the care plan is a generative task. The prediction needs a well-defined outcome and calibration check; the drafted summary needs strict fidelity to the clinician’s orders. Combining them in one interface does not remove either requirement. Describe the output you need before deciding what evidence would make it useful.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy