What is ground truth?
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.
Ground truth is the reference answer or observation used to judge a system’s output. It establishes what the evaluation treats as correct. The name describes its role as a comparison standard; it does not mean that every reference record is complete, unquestionable or free from error.
For handwritten-digit recognition, a reference label might identify the digit in each picture. For a delivery forecast, the reference may be the recorded arrival time. For an invoice-reading task, it might be the supplier and total checked by an experienced employee. Prior expert work can provide a reference, but it may still need checking.
Different tasks need different kinds of reference. Some have directly observable answers. Others involve judgment, such as whether a summary preserves the important qualifications in a report. In those cases, the evaluation needs a clear description of what reviewers should assess and how disagreements are handled.
A label is an answer attached to an example. Labels can be mistaken, inconsistent or based on information unavailable to the system at the relevant moment. A reviewer may also infer something that the source itself does not establish. Calling the label ground truth should not prevent it from being examined.
The reference should match the intended question. A recorded customer complaint is not the same as every instance of customer dissatisfaction. A historical decision is not necessarily the decision that should have been made. These records may be useful, but their meaning must be stated rather than silently expanded.
For generated text, which may contain several claims, there may be no single correct sentence. A useful reference can specify required facts, acceptable variations and unsupported additions. It should distinguish a different wording from a genuinely different conclusion. Agreement with one preferred phrasing is not always the right standard.
Ground truth also needs maintenance when facts change. A current policy answer should not be judged against a retired policy simply because the older answer is already stored in a test collection.
The psychiatric care-pathway question makes the distinction especially important. The setting a patient actually entered is an observable outcome; it is not automatically ground truth for the most appropriate setting. A research team might instead need a defined clinical-review process that allows multiple acceptable options and uncertainty, while separately recording service availability. That is a proposed way to construct a reference, not a result established by the class discussion. Before trusting a score, ask exactly what its reference answer represents.
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