Learn AI /How Machines Learn /Ground truth
Ground truth
- Intermediate
- How Machines Learn
- Technical readers
- Programme & data teams
- Human rights practitioners
Also called: gold standard.
Back to the milestone you came from
In five seconds
What it is
The checked reference answer used to train or evaluate a model, so its output can be compared with the best available real-world answer.
In plain words
Ground truth is the answer you use as the trusted reference when checking whether a model got something right.
An everyday example
A class in Nairobi, Bogotá and Bangkok is reviewing short sentences that contain the word “she”. For each sentence, two language teachers read the surrounding text and mark who “she” refers to, such as Amina or Sofia. Their agreed answer is stored in a spreadsheet. Later, a language tool reads the same sentences and its answer is checked against that agreed reference.
Real machine-learning ground truth is often larger and less tidy than this story. Experts may not fully agree, some tasks use measured numbers rather than names, and training usually uses many examples with mathematical updates, not a simple one-by-one classroom check. Also, in some domains there may be no single unquestioned answer, only the best available annotation under agreed rules.
Remember thisGround truth is the checked reference target used to train or evaluate a model. It is not the model's answer and, despite the name, can still reflect judgement, ambiguity or labelling error.
This term is contested
The core idea is widely used, but the term can be contentious because 'truth' may overstate certainty in domains where reference answers are partial, uncertain or socially contested.
Understand it
Understand it
Why it matters
If you hear that a model is “90% accurate”, the next question should be: compared with what reference answer? That reference is the ground truth. If it is weak, outdated, biased or inconsistent, the performance claim can mislead decision-makers. Ground truth affects procurement, risk assessment, quality assurance and fairness review. It helps you ask practical questions: Who created the reference answers? Were clear guidelines used? Did experts agree? Does the reference reflect the population and language varieties the system will face? A non-specialist does not need to build models to see that better reference answers usually lead to more trustworthy evaluation and, often, better training.
How ground truth is used in training and testing
After viewing this, the reader should be able to say that ground truth is the reference answer used to compare a model’s output during training or evaluation.
01
1. Example enters
Start with one input, such as a sentence, image or record.
Draw a rectangle on the left labelled 'Input example'. Show small icons for text, image and table beneath the label.
02
2. Model produces output
The model turns the input into a prediction, such as a class, number or extracted span of text.
From the input box, draw an arrow to an upper rectangle labelled 'Model', then an arrow to a box labelled 'Prediction'.
03
3. Reference answer is provided
Separately, the same input has a human-checked or otherwise trusted reference answer.
From the input box, draw a second arrow downward to a lower box labelled 'Ground truth reference' with a small checkmark icon.
04
4. Comparison happens
The prediction is compared with the reference answer to measure mismatch or agreement.
Draw arrows from 'Prediction' and 'Ground truth reference' into a central comparison box labelled 'Compare'.
05
5. Output of the comparison
In training, the mismatch helps update the model. In evaluation, it contributes to metrics such as accuracy or error.
From 'Compare', split into two right-side boxes: upper labelled 'Training update' and lower labelled 'Evaluation score'.
A left-to-right flow diagram with two lanes. Both lanes start with the same kind of input example. In the top lane, the example goes to a model and produces a prediction. In the bottom lane, the same example links to a human-checked reference answer. The two meet in a comparison stage that outputs either an update signal for training or a performance score for evaluation.
Hovering or tapping each box highlights the arrows connected to it and reveals one sentence explaining its role.
Describe this without the diagram
The diagram is a two-lane flow showing how ground truth is used. First, an input example enters. In the top lane, the input goes through a model and becomes a prediction. In the bottom lane, the same input connects to a ground-truth reference answer that has been checked by people or another trusted process. The prediction and the reference answer then meet in a comparison step. The result of that comparison has two possible uses: during training, the mismatch helps update the model; during evaluation, the comparison contributes to a performance score such as accuracy or error. The key idea is that ground truth is the reference used for comparison, not the model’s output and not the metric itself.
With reduced motion: A static two-lane flowchart shows one input feeding both a model path and a reference-answer path. These meet in a comparison box, which then points to either training updates or evaluation scores.
In machine learning, ground truth means the reference answer you treat as the best available answer for a case. It might be a correct category, a measured number, a transcript reviewed by people, or a location marked on an image. People also say “gold standard”, although in practice it may be imperfect.
What goes in is an example and its reference answer. During training, a model processes many examples and tries to produce outputs close to those reference answers. During evaluation, the model’s output is compared with the reference answer to see how well it performed.
What happens inside depends on the task. In a classification task, the model may compare its predicted class with the ground-truth class. In a regression task, it may compare a predicted number with a measured number. The difference between the model output and the reference answer becomes an error signal, which helps measure performance and often helps adjust the model during training.
A label and ground truth can contain the same value, but the terms describe different roles. A label is the expected answer attached to an example. It serves as ground truth when it is treated as the trusted reference for comparison. Ground truth can also come from a measurement or another reference process. Accuracy is different: it is a summary score calculated after comparing many predictions with ground truth.
A practical caution: ground truth can be noisy, biased, delayed or disputed. For example, several experts may disagree, or the “true” answer may change over time. So ground truth is usually the best available reference, not a guarantee of perfect truth.
How “Checking who “she” refers to in a sentence” maps onto the real thing
| In the story | In the AI system |
|---|---|
| The sentence with an ambiguous pronoun | The input example |
| The teachers’ agreed answer about who “she” refers to | The ground-truth reference answer |
| The language tool’s guess | The model prediction |
| Comparing the guess with the spreadsheet entry | Training or evaluation against ground truth |
A worked example
A service desk focal point at Headquarters is evaluating a classifier before figures go into a monthly monitoring report. Reviewers have labelled a held-out set of requests using an agreed coding guide.
Starting from: Sanitised request text: "I cannot log in after resetting my password."
- Reviewers check the request and assign the category "account access". This checked label is the ground truth for the example.
- The model assigns probabilities of 0.82 to "account access", 0.12 to "software request" and 0.06 to "other".
- For evaluation, the model's highest-probability class is compared with the ground-truth class. They match, so the prediction is counted as correct.
- Because this request is in a held-out evaluation set, the result contributes to performance metrics but does not update the model.
- If the labelled request were used during training, a loss function would compare the predicted probabilities with the ground-truth target and use the difference to adjust model parameters.
Result: The ground truth is "account access". The classifier receives credit for a correct prediction, while its probability distribution still shows how much uncertainty it assigned to the alternatives.
In a United Nations workplace
At close of business, a human resources officer in a Resident Coordinator's Office is checking the staffing section of a quarterly report to a donor. A pilot tool has classified recruitment actions as on time or delayed. The officer compares its outputs with labels assigned by reviewing the approved workflow timestamps against agreed reporting rules. Those checked labels are the ground truth.
If the reference labels use inconsistent cut-off rules or omit exceptional cases, a high accuracy score says little. The officer needs to record who assigned the labels, how disputed cases were resolved and whether the held-out sample reflects the recruitment actions covered by the report.
Before you try thisRecruitment files contain personal and HR-sensitive information. They must not be uploaded to a tool that has not been approved for such material. Use the approved environment, remove fields not needed for classification, restrict access and retain a human check before the figures are cleared.
Where you may hear it
- A vendor demo during a procurement meeting
- A quality assurance meeting before an evaluation report is cleared
- An ICT governance board reviewing a pilot
- A labelling review with subject-matter focal points
Avoid confusion
Avoid confusion
Often confused with
Label
The expected answer attached to an example so a model can learn what output should match that input.
Prediction
The result a model produces from input, such as a class, number, ranking or next token.
Accuracy
The share of cases a system labels correctly out of all the cases it was tested on.
Common misconceptions
Common misconception: Ground truth means absolute truth with no uncertainty.
Correction: Often it means the best available reference under a chosen method, not perfect certainty. In medicine, language, human rights documentation and other complex areas, experts may disagree or the answer may be incomplete. Good practice is to document how the reference was produced and how uncertainty was handled.
Common misconception: Ground truth is just another word for label.
Correction: A label is one kind of reference answer, usually a tag such as 'spam' or 'not spam'. Ground truth is broader. It can be a label, a measured number, a bounding box in an image, a transcript, or another trusted target used for comparison.
Common misconception: If the ground truth is human-made, it cannot be biased.
Correction: Human-created references can reflect inconsistent rules, limited expertise, cultural assumptions or skewed sampling. A model trained on such data may reproduce those patterns. That is why annotation guidelines, reviewer agreement and dataset coverage matter.
Common misconception: High accuracy proves the model works well everywhere.
Correction: Accuracy only summarises performance against a particular ground-truth set. If that set is narrow, outdated or unlike real deployment conditions, the score may not transfer. You need to examine the reference data, not only the final metric.
Quick check
Which statement best describes ground truth in machine learning?
Ground truth is the trusted reference answer for an example. A model prediction is compared with it. Accuracy is a metric calculated after many such comparisons, and features are parts of the input, not the reference answer.
Go deeper
Go deeper
The simple version
Ground truth is the answer you use as the trusted reference when checking whether a model got something right.
The technical version
Ground truth is the target value or reference annotation assigned to an example and treated as the benchmark for comparison during training, validation or testing. In supervised learning, the model receives inputs and associated targets, then a loss function measures the difference between predicted outputs and ground-truth targets. Those differences can be used to update model parameters during training or to compute evaluation metrics on held-out data. The reference may come from human annotation, direct measurement, sensor readings, administrative records or expert adjudication. Depending on the task, it may be categorical, numeric, textual or spatial. A label is an expected answer attached to an example; it is ground truth when treated as the trusted reference. Ground truth is also distinct from prediction, which is the model output, and from accuracy, which summarises comparisons across many examples. In practice, ground truth may be noisy, partial, delayed or contested, so dataset documentation and annotation protocols are part of technical quality control.
Reviewed external resources
EncyclopaediaGround truth (opens in a new tab)Wikipedia
This link gives a broad overview of the term and its uses across fields, which suits readers who want a general starting point.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course resource helps readers see how reference answers are used in practical machine-learning workflows, which suits beginners moving into applied work.
Intermediatelink checked 2026-07-31
VideoStatQuest (opens in a new tab)StatQuest with Josh Starmer
This video source is useful for readers who learn well from spoken step-by-step explanations of model training ideas.
Starterlink checked 2026-07-31
VideoVideo explanations: Ground truth (opens in a new tab)YouTube search
This is a YouTube search results page, so it helps readers browse multiple video explanations and choose a style that suits them.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Ground truth (opens in a new tab)Reddit search
This is a Reddit search results page, so it helps readers see how practitioners informally discuss the term, though comments should be checked carefully.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Ground truth is the checked reference target used to train or evaluate a model. It is not the model's answer and, despite the name, can still reflect judgement, ambiguity or labelling error.