Definition
Ground truth
Ground truth is the verified body of knowledge an answer is checked against: the documents an organization has decided are correct and authoritative. It is the thing an AI answer has to be grounded in, and the thing you audit against.
Ground truth is the verified body of knowledge an answer is measured against: the drawings, specifications, regulations, and records an organization has decided are correct and authoritative. It is what “grounded” refers back to. An answer is only as reliable as the ground truth it was grounded in, and only as auditable as that ground truth is defined.
How is ground truth different from an AI model’s training data?
Training data is the vast, mixed corpus a model learned language from. It is broad, unverified, and not yours. Ground truth is narrow, verified, and yours: the specific documents your organization stands behind. A generalist model answers from its training data, which is why it can be fluent about your domain and wrong about your specifics at the same time. A grounded answer is checked against your ground truth, which is the body of knowledge you actually trust. The distinction is the whole reason grounding exists.
Why is ground truth the thing you audit against?
Because reliability is not a property of the model. It is a property of the answer measured against a known correct source. If there is no defined ground truth, “reliable” has nothing to mean, because there is nothing to be reliable with respect to. Defining the ground truth is what makes it possible to say whether an answer is right, to catch it when it is wrong, and to prove sufficiency when the cost of a missed clause is operational. The claims rule the grounding service holds itself to depends on it: answer from the ground truth, or name what is missing from it, never guess past its edges.
What happens at the edge of the ground truth?
The grounding service treats the edge of the ground truth as a hard boundary. If a question cannot be answered from the verified sources, the service says so and names what it would need, rather than reaching past the boundary for a better-sounding answer. An answer that quietly drifts outside the ground truth to satisfy the question is exactly the failure that makes an AI untrustworthy in the work that matters. Holding the boundary is what keeps every answer accountable to a source you have verified.
Related terms
- Grounding: answering from the ground truth instead of guessing.
- Provenance: the origin and revision that make a source part of the ground truth.
- Hallucination: what happens when an answer leaves the ground truth behind.