Ground truth is the set of labels or data accepted as objectively correct, used as the standard a model's predictions or an annotator's work gets measured against. It's not necessarily a perfect, universal truth, it's the best available correct answer a team has agreed to trust.

‍

‍Why it matters

Every accuracy number, every QA check, every model evaluation depends on having something to compare against, and ground truth is that reference point. The word truth can be a little misleading though, ground truth is still created by people making judgment calls or by processes that can themselves contain errors, which is why establishing it carefully, often through a gold standard review process, matters so much. A model can only ever be as good as the ground truth it's measured against, if the ground truth itself is flawed, a model that matches it perfectly is still wrong in the same ways.

‍

‍Where it comes from

Ground truth is typically established through expert-verified gold standard datasets, direct observation or measurement where possible, or thorough consensus review among multiple qualified annotators, rather than accepted from a single source without verification. Because it carries so much weight, ground truth data usually goes through a stricter review process than regular labeled data, since an error here doesn't just affect one item, it distorts every accuracy measurement that relies on it afterward.