Skip to main content

What is AI transparency?

The 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.

AI transparency means making relevant information about a system available and understandable to the people who need it. It can concern how the system was developed, what it is intended to do, what information it uses, what actions it has taken and what its limitations are.

Transparency is not the same as publishing every technical detail. A patient may need to know that a message was drafted with AI and how to question it. A clinician may need the evidence supporting a recommendation. An engineer investigating a failure may need a version identifier and a record of tool calls. These are different information needs.

Model cards and dataset documentation can help describe intended uses, development data, evaluation and limitations. A model card is a structured report about a model; a dataset document describes the information used. Their value depends on the quality and specificity of the content, not simply the presence of a template.

Transparency also differs from explainability. Explaining why an output occurred is one possible form of transparency. Disclosing that the system was not evaluated in the intended population is another, and may be more important for a deployment decision.

In an agentic application, a useful record distinguishes a proposed action from an executed action. A generated plan is not an audit log. An audit log records relevant events, such as the request, permission decision, tool response and resulting change. Even that record needs appropriate access controls because it may contain sensitive information.

For a referral assistant, staff should be able to see which policy and patient document supported a request for additional information. A patient should not need to understand the model architecture to challenge an incorrect requirement. Some regulations and certification programs impose specific disclosure obligations, but those obligations vary. Transparency should therefore be designed around both applicable requirements and practical accountability. Giving people more information is helpful only when it allows them to understand, check or contest something that matters.

Sources: Mitchell and colleagues — Model Cards for Model Reporting (2019); Gebru and colleagues — Datasheets for Datasets (2021); ASTP/ONC — Decision Support Interventions Certification Companion Guide

To learn more, take a look at 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

Lead with AI in Health Care

The AI Pragmatist is a monthly newsletter from ALA faculty, with new explainers & essays from the Advanced Learning Academy