Ground Truth is the correct, verified answer for a piece of data — the “reality” against which a model’s predictions are compared to measure how accurate it is.
What it means in plain English
To know whether a model is right, you need to know the true answer. Ground truth is that trusted reference: the actual, correct label or outcome. During training, ground truth tells the model what it should have predicted; during evaluation, it is the standard the model’s output is judged against. The quality of your ground truth sets a ceiling on how good and how well-measured your model can be.
Obtaining reliable ground truth can be difficult, expensive, or occasionally ambiguous.
A simple example
To evaluate a medical AI, doctors provide the confirmed diagnosis for each case — the ground truth. The model’s predictions are compared against these verified answers to measure how often it is correct.
Why it matters
Ground truth is the foundation of both training and honestly evaluating AI. Without accurate ground truth, you cannot properly teach a model or trust any measurement of its performance — making it one of the most important, if unglamorous, ingredients in machine learning.
Related terms
- Data Labelling — how ground truth is often created.
- Benchmark — uses ground truth to score models.
- Training Data — includes the ground-truth answers.
Frequently asked questions
What is ground truth?
Ground truth is the verified, correct answer used to train and evaluate a model — the reference “reality” its predictions are compared against.
Why does ground truth quality matter?
If the ground truth is wrong or inconsistent, the model learns from and is judged against bad references, so both training and evaluation suffer. Accurate ground truth is essential.
