What is an AI model?
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.
An AI model is the mathematical component of an AI system that turns information it receives into a prediction, representation or generated response. Most models discussed today learn patterns from examples. They can then apply those patterns to new information, rather than relying only on rules that a programmer has written individually.
A simple example is a system that predicts the sale price of a house. It might learn relationships between price, location, size and other characteristics. The characteristics supplied to the model are called features. The predicted price is its output: the result it produces from the information it receives.
Training is the process of adjusting the model using examples. In many models, this changes internal numbers called parameters. These numbers determine how strongly different patterns influence the result. A simple model may have relatively few parameters; a large language model, which processes and generates language, can have many more.
Using the trained model to make a prediction or generate an answer is called inference. Training and inference are different activities. A model does not necessarily learn permanently from every question it answers. An application may remember your conversation, but saving that conversation is not the same as changing the model’s parameters.
The model is also not the whole product. A chatbot application may combine a model with document search, a calculator, stored conversation history and rules about what actions are allowed. Two products using the same model can behave differently because they supply different information and connect it to different functions.
Nor is a model simply a database: an organized collection of stored records that can be retrieved. Its learned parameters capture patterns, but they do not provide a dependable record of every source or fact encountered during training. Some applications therefore connect models to databases when they need current or verifiable information.
A psychiatric care-pathway question helps separate a model from a clinical service. A model might estimate an outcome from information available at intake. It would not, by itself, establish which services are available, what the patient prefers or who is responsible for arranging care. Those belong to the surrounding clinical and operational process. Start by specifying the model’s input, intended output and evidence, then ask what additional work is needed before that output can help a patient.
Sources: Google — What Is Machine Learning?; Mitchell and colleagues — Model Cards for Model Reporting.
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