Learn AI /Data and Infrastructure /Sampling (of data)
Sampling (of data)
- Starter
- Data and Infrastructure
- Programme & data teams
- Human rights practitioners
Also called: sample selection, representative sample.
Back to the milestone you came from
In five seconds
What it is
Choosing a smaller set of cases from a larger group so you can measure, inspect or train on it without using every case.
In plain words
Sampling means taking a smaller, manageable slice of a bigger set of data so you can learn something about the bigger set.
An everyday example
A shop receives many boxes of fruit and cannot inspect every piece before opening. Checking only the first box may miss damage elsewhere in the delivery. Taking fruit from several boxes, including boxes from different parts of the stack, gives a better picture. The conclusion still depends on how those pieces were chosen and whether every part of the delivery had a fair chance of being checked.
The story suggests people can deliberately balance the sample by eye, but many data tasks need stricter methods, such as random sampling, to reduce hidden bias. Also, in machine learning, the sample may contain millions of rows or files, and the goal may be training a model rather than making a one-time estimate. Finally, some projects do not want a sample to represent the whole population; they may intentionally over-sample a rare but important group.
Remember thisA sample is useful only in relation to how it was selected. The selection method shapes bias, uncertainty and how far the findings can be generalised.
Understand it
Understand it
Why it matters
Non-specialists often see a chart, dashboard or model result and assume it reflects reality. But if the underlying sample was poorly chosen, the result can be misleading even when the maths looks precise. This affects programme planning, human-rights analysis, procurement, monitoring and evaluation, and model training. A sample that misses rural areas, minority languages, rare harms or low-connectivity users can lead to unfair conclusions and weak decisions. Asking 'How was the sample chosen?' is often more important than asking only 'How large was it?'
How a sample is chosen from a larger data source
After viewing this, a reader should be able to explain that sampling means selecting a smaller subset from a larger group, and that the selection method affects whether results can be trusted.
01
1. Start with the full group
Show a large collection of cases such as people, records or documents. This is the source you want to learn about.
Draw a large container with many small dots or cards inside. Add a label: 'Full group / source data'. Use mixed colours or shapes to suggest variety across regions or categories.
02
2. Choose a sampling method
The method might be random, stratified by group, or convenience-based. The method matters because it affects who gets included.
Show three small selection tools above the container: a dice icon for random, a grouped grid for stratified, and a hand reaching to the nearest items for convenience. Highlight one active path.
03
3. Create the sample
A smaller subset is pulled out. If important groups are missing or over-represented, the sample may not reflect the full group well.
Draw an arrow from the large container to a smaller container holding fewer dots or cards. Include a note such as 'smaller subset' and show whether the mix of colours roughly matches the original.
04
4. Analyse and interpret carefully
Results from the sample are used to estimate, inspect or train, but confidence depends on how the sample was chosen.
Draw a final panel with a simple chart or checklist fed by the sample container. Add two short tags: 'useful if representative' and 'risky if skewed'.
A left-to-right step diagram shows a large pool of data cases, the choice method, the resulting sample, and the conclusion drawn from it. Read from first step to last step to see how selection choices shape what you can say about the whole group.
On hover or tap, each step reveals a short note with one risk: missing groups, convenience bias, small sample size, or overconfidence in conclusions.
Describe this without the diagram
Step 1 shows a large and varied source group of data cases. Step 2 shows that there are different ways to choose cases, such as random selection, balancing by groups, or simply taking the easiest cases to reach. Step 3 shows the smaller sample that results from that choice. If the sample still contains the important variety of the full group, it can support more reliable analysis. If some groups are missing or over-represented, conclusions may be distorted. Step 4 shows that any estimate, audit or model built from the sample should be judged in light of the selection method, not only the final numbers.
With reduced motion: A static four-panel diagram appears with the same sequence: full group, choice method, sample, and conclusion. Notes under each panel explain the main risk without animation.
In data work, sampling means selecting some records, people, documents or events from a larger group. The goal is to learn something useful about the whole group without checking every single item. The selected subset is called a sample.
What goes in is a larger source of data, such as all survey responses, all clinic records approved for analysis, or all images in a collection. What happens is that someone chooses part of that larger set, using a method such as random selection or deliberate balancing across groups. What comes out is a smaller set that will be analysed, checked or used for training.
A good sample reflects the larger group closely enough for the task. People often call this representative. For example, if a population includes different regions, languages or age groups, the sample should not accidentally leave some of them out unless that is an intentional and justified choice.
This is often confused with data distribution and with sampling in text generation. Data distribution describes how values are spread in the data, such as whether most cases come from one country or one class. Token sampling is the process of choosing the next word or symbol from a model's output probabilities. Here, sampling means choosing data cases from a larger pool.
How “Checking a delivery of fruit” maps onto the real thing
| In the story | In the AI system |
|---|---|
| households interviewed | sampled records or cases |
| all households using the distribution point | full population or source dataset |
| including different family sizes, times and areas | trying to make the sample representative |
| estimate for next week's supplies | analysis or model built from the sample |
A worked example
An information management officer at a regional hub must check the consistency of feedback records before a monitoring report is circulated, but reviewing the entire dataset would miss the deadline.
Starting from: The source dataset contains records submitted through several channels and reporting periods, with some groups much larger than others.
- The officer defines the target population and the fields to be checked before drawing any records.
- Rather than reviewing the easiest records to retrieve, the officer divides the dataset into groups by submission channel and reporting period.
- Records are selected randomly within each group, with the number selected reflecting the purpose of the audit and the size of each group.
- The officer documents the method and checks whether missing records, duplicate entries or unequal response rates could still bias the findings.
Result: The sample provides a defensible basis for assessing data quality across the source dataset. The monitoring report records how the sample was drawn and avoids stronger claims about groups for which coverage remains limited.
In a United Nations workplace
A national officer in a liaison office is finalising a statement for delivery that cites findings from partner feedback. Before sending it for clearance, she checks whether the forms analysed were sampled across submission channels and reporting periods, rather than simply taken from the top of the file.
The sampling method determines how confidently the statement can describe the wider body of feedback. Random selection within relevant groups can reduce the risk that one channel or period dominates, while a convenience sample may support only a narrower claim.
Before you try thisFeedback forms may contain personal or otherwise sensitive information. Do not upload forms or identifying details to a tool that has not been approved for that material. Use only authorised systems, minimise the data accessed, and state any limits in coverage or representativeness before the finding is cleared.
Where you may hear it
- A clearance discussion on figures cited in a statement
- An evaluation office meeting on survey design
- A vendor demo on selecting model training records
- An internal audit opening meeting on record quality
- An inter-agency working group reviewing a needs assessment
Avoid confusion
Avoid confusion
Often confused with
Data distribution
The pattern of values, categories and frequencies in a dataset, including what appears often, rarely or not at all.
Sampling (choosing the next token)
A stochastic procedure that draws one next token from the model’s probability distribution, appends it to the output, and repeats until a stop condition.
Common misconceptions
Common misconception: A large sample is automatically a good sample.
Correction: Size helps, but it does not fix bad selection. A very large convenience sample can still miss key groups or over-represent easy-to-reach cases. A smaller but well-designed sample may support better conclusions.
Common misconception: Random sampling means every result will perfectly match the full population.
Correction: Random selection reduces systematic bias, but chance variation still exists. A random sample can differ from the full group, especially if the sample is small or the population is very diverse.
Common misconception: If a sample is not perfectly representative, it is useless.
Correction: Not necessarily. A sample can still be useful for some tasks, such as spotting obvious quality problems or testing a process. The key is to match the sampling method to the purpose and to state the limits clearly.
Quick check
Which choice gives the best basis for reviewing whether service complaints are similar across four regions?
Random selection within regions and proportional allocation preserve the regional make-up while reducing selection bias. Equal weighting changes that make-up, and selecting recent or easy-to-access cases may skew the review.
Go deeper
Go deeper
The simple version
Sampling means taking a smaller, manageable slice of a bigger set of data so you can learn something about the bigger set.
The technical version
Sampling is the procedure used to select observations from a target population or source dataset for analysis, auditing or model development. Selection may be probability-based, such as simple random sampling or random sampling within defined strata, or non-probability-based, such as convenience sampling. The method affects representativeness, sampling bias, variance and how confidently findings can be generalised. In machine learning, the sampled dataset shapes what patterns the model can learn, because under-sampled groups or events may be poorly represented in training or evaluation.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
A practical official framework for managing AI risk, useful for staff who need to connect data choices such as sampling to governance and oversight.
Intermediatelink checked 2026-07-31
Primary sourceOECD AI Principles (opens in a new tab)OECD
A high-level official principles document, useful for readers who want to place careful data practices within wider responsible AI expectations.
Intermediatelink checked 2026-07-31
EncyclopaediaSampling (statistics) (opens in a new tab)Wikipedia
A broad introductory reference on sampling in statistics, useful for readers who want definitions, common methods and background terms.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Sampling (of data) (opens in a new tab)YouTube search
A YouTube search results page, useful for readers who prefer short video explanations and want to compare several teaching styles before choosing one.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Sampling (of data) (opens in a new tab)Reddit search
A Reddit search results page, useful for readers who want to see common questions and informal discussion, but it should not be treated as an authoritative source.
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
A sample is useful only in relation to how it was selected. The selection method shapes bias, uncertainty and how far the findings can be generalised.