1. Home
  2. Glossary
  3. Ground truth
AI glossary · Models & training

Ground truth

Ground truth: Ground truth is the verified, correct answer that an AI system's output is compared against, such as a human-labeled dataset or a confirmed real-world outcome. It is the reference standard used to train models and measure their accuracy.

Every claim about an AI model's accuracy depends on something being treated as correct. That something is the ground truth. For a model that classifies support tickets, it is a set of tickets that experienced agents categorized by hand. For a demand forecast, it is what actually sold. For a document extraction tool, it is the fields a person checked against the original.

Ground truth is expensive because it usually requires people. Someone has to label the examples, and those people disagree, get tired, and bring their own assumptions. When ground truth is wrong or inconsistent, the model learns the wrong thing and the accuracy score is meaningless, because you are grading against a flawed answer key. Good teams measure how often their labelers agree with each other before they trust the labels.

In generative AI, ground truth is harder to pin down. There is no single correct summary of a meeting or one right way to write an email. Teams handle this by having experts rate outputs, by writing rubrics, or by checking for specific facts an output must contain. When a vendor claims a model is accurate, ask what the ground truth was, who created it, and how similar it is to your own data.

Example at work

A medical coding team pilots an AI tool that suggests billing codes from clinical notes. After confirming the tool is approved for PHI, and before trusting the vendor's accuracy claims, they pull two hundred past encounters their senior coder has already audited and treat those codes as ground truth. The tool's accuracy on that set, not the vendor's marketing figure, decides whether the pilot expands.

Why it matters

Whenever someone shows you an accuracy number, ground truth is the thing that number was measured against. Asking where it came from is the fastest way to tell a meaningful evaluation from a hollow one, and to spot when a model was graded against data that looks nothing like yours.

Related terms