What are permissions in an AI application?
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.
Permissions determine which information an application can access and which actions it may perform. In an AI agent, they help separate work that may happen automatically from work that requires approval or is unavailable altogether. An agent is software that chooses actions while pursuing a goal; permissions limit what those choices can actually cause.
Consider the difference between reading, drafting and sending email. Reading an email reveals information. Drafting prepares a possible response. Sending communicates with another person. These are different actions with different consequences, even when they occur through the same familiar interface.
Authentication and authorization answer related but separate questions. Authentication checks identity: which user or application is making the request? Authorization determines what that identity may do. A valid sign-in does not necessarily grant access to every document or permission to send every message.
Permissions can apply at several levels. A connected service may restrict which account or folder is accessible. A tool may expose only particular operations. The application may require approval for a specific proposed change. Clear design makes these levels work together rather than assuming one login settles everything.
A prompt saying “do not send email” is not equivalent to removing the sending capability. The prompt is an instruction for the model, the trained component proposing actions. A software-enforced permission check can block a prohibited operation regardless of the wording the model generates. Instructions and enforcement should support one another.
Approval also needs a clear scope. Agreeing to send one draft to one recipient does not automatically authorize later drafts, changed attachments or a larger recipient list. The user should be able to understand what action is about to happen and what information it will expose.
A referral-support agent might be allowed to read an episode’s administrative status and prepare an internal task, but not edit clinical notes or send a patient message. Even an approved message needs a defined recipient and content: changing either can change what was authorized. The goal is enough access for the assigned work, not access to everything the organization knows. In a clinical organization, useful permissions make the difference between reading, preparing, approving and acting explicit.
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