How should model and harness selection be understood?
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.
Model selection and harness selection are two related decisions. The model is the trained component that interprets information and proposes responses. The harness is the surrounding software that supplies relevant material, offers tools, manages access and carries work through repeated steps. Choosing one does not automatically settle the other.
Choose a model you can afford and rely on for the task, in an environment you can work with effectively. “Trust” should be made concrete: can you inspect the work, recognize a mistake, correct the result and repeat the task? A familiar brand or impressive demonstration is not enough to answer those questions.
Start with a task you understand. For example, ask each candidate setup to summarize the same short document while preserving its main qualifications. Establish a reference account of what a good answer must contain. Then examine what the system produces, how much correction it requires and whether the result is repeatable enough for the purpose.
The comparison should distinguish components where possible. Changing the model while keeping the instructions and documents similar helps identify the effect of the model. Changing the environment may also change tools, available context and review controls. Context is the information supplied for the task. A better result may come from better context rather than a stronger model.
Cost includes more than one request. Repeated attempts, checking and correction can change the total effort substantially. A cheaper response that regularly needs rescue may not be cheaper work. A more expensive model may also be unnecessary for a simple task that ordinary software can perform exactly.
The harness should support the level of control the task requires. Can you distinguish reading from editing? Can you see what actually ran? Can you stop or recover the work? These are operational capabilities, not measures of how fluent the model sounds.
For a clinic handoff assistant, compare candidate setups on the same fictional referral cases, including one with a missing document and one with conflicting status records. Judge whether the brief identifies the right next administrative step, preserves uncertainty and lets staff inspect the evidence. Also check what actions the harness permits and how much correction is needed. A better-sounding model response is not necessarily a better working arrangement. Choose against a small task the team understands, rather than a general ranking.
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