What is generative AI?
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.
Generative AI is artificial intelligence that produces content, such as text, images, sound, video or computer code. It learns patterns from examples and uses those patterns to create an output in response to instructions or other information. The defining feature is generation: producing content rather than only sorting or scoring existing material.
For example, a system might write a birthday invitation from a description, create an illustration for it or generate accompanying music. A different AI system might only identify which photographs contain a birthday cake. Both use artificial intelligence, but creating the invitation and recognizing the cake are different tasks.
The component that learns these patterns is called a model. During training, the model’s internal numerical settings are adjusted using examples. Once trained, it can produce outputs based on what it receives. Your instruction is often called a prompt. A prompt could be written text, an image or a combination of information.
Different kinds of generative AI work differently. Many text systems generate small pieces of language one after another. Some image systems use diffusion: they learn to reverse a process that gradually adds random visual noise to images. When generating, they start from noise and progressively shape it into a picture, often guided by a description.
Generation is not the same as looking up a stored answer. A system can create a sentence that never appeared in its training examples. However, it can also reproduce or closely resemble existing material. “Generated” does not automatically mean original, accurate or free from restrictions on how the material may be used.
This distinction is especially important with factual questions. A fluent answer can include an invented event, quotation or source. These unsupported outputs are often called hallucinations. The term describes a failure of accuracy; it does not imply that the computer has a human experience.
A clinical team might use generative AI to turn an already agreed care plan into a draft patient-facing explanation. The useful transformation is clearer language, not a new diagnosis, altered medication instruction or invented follow-up. The clinician would need to check the draft against the source plan, and the team would need to test whether patients understand it. Generating a plausible explanation and faithfully communicating an agreed clinical decision are different achievements.
Sources: NIST — Generative Artificial Intelligence Profile; Ho and colleagues — Denoising Diffusion Probabilistic Models.
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