Suppose a bank builds a model to flag fraudulent card payments. To know whether it works, the team needs a set of past payments where the outcome is already known: these were confirmed fraud, these were legitimate. Those confirmed outcomes are the ground truth. The model’s guesses are scored against them, and the same kind of labeled examples are used to train it in the first place (see supervised learning).
Ground truth usually comes from people or from what happened later: an expert’s diagnosis, a customer’s confirmed chargeback, a test that passes. Creating it is the job of data labeling. It is only as good as the process behind it. Labelers disagree, mistakes creep in, and a “correct answer” set can go stale as the world changes.
Why it matters: generative AI makes ground truth harder to pin down. A support reply or a summary can be good in many different ways, so there is often no single right answer to compare against. Teams fill the gap with human review, curated “golden” test sets, rubrics, LLM-as-a-judge scoring or real-world outcomes. Each is a stand-in, and knowing which one an evaluation uses tells you how much to trust its score.
In episode 37 at 18:50, Galileo co-founder Atindriyo Sanyal calls the habit of “comparing an output with a pre generated ground truth, which is either human labeled or synthetic” a core part of data literacy. In episode 49 at 18:38, Intercom’s Fergal Reid explains that offline tests are not enough; confirmed customer resolutions in production are “our ultimate ground truth.” In episode 4 at 10:00, Galileo CEO Vikram Chatterji lists ground truth among the pieces teams need to evaluate an AI system. Sponsor note: Galileo sponsored the show when these episodes aired. Sanyal is introduced as Galileo’s CTO in the episode 37 opening, Chatterji’s Galileo role is introduced in the episode 4 opening, and episode 49 includes a Galileo sponsor read.
“Ground truth” is also everyday English for “what really happened,” so this glossary does not link it automatically.
Hear it from the guest
“Comparing an output with a pre generated ground truth, which is either human labeled or synthetic”
“Our ultimate ground truth”
“You also need the ground truth”
Quotes lightly edited to remove filler words.
Sources
- Google Machine Learning Glossary: Ground truth — Defines ground truth as reality or what actually happened, distinguishing observed outcomes from a model's prediction; checked October 24, 2026.
- Ratner et al.: Snorkel: Rapid Training Data Creation with Weak Supervision — Describes labeling training data as a major bottleneck and a system that combines noisy labeling rules without access to ground truth; checked October 24, 2026.
