Learn AI /Data and Infrastructure /Synthetic data
Synthetic data
- Intermediate
- Data and Infrastructure
- Programme & data teams
- Policy officers
- Technical readers
Also called: artificial data.
Back to the milestone you came from
In five seconds
What it is
Artificially generated records designed to resemble data from a real population or process for uses such as testing, training or sharing.
In plain words
Synthetic data is artificially created data that resembles real data closely enough for practice, testing or some kinds of training.
An everyday example
A training officer wants new colleagues to practise summarising meeting notes, but the real notes include names, sensitive topics and unfinished decisions. She studies several real note sets to see the usual structure, common agenda items and the way action points are written. Then she creates many new, fictional note sets that follow similar patterns but do not describe actual people or actual meetings. The trainees use those practice notes to learn the task before they handle approved real material.
In the story, a person intentionally writes each practice note. In real synthetic data, software often generates very large volumes automatically. Also, fictional meeting notes may avoid all real details, while synthetic data can still leak information if the method copies too closely from the source. Finally, some real-world patterns are much harder to recreate than meeting-note structure, especially rare events and hidden relationships.
Remember thisSynthetic data can support testing, training and sharing, but it is not automatically anonymous. Its usefulness and disclosure risk must both be measured for the intended task.
Understand it
Understand it
Why it matters
Synthetic data affects practical decisions about access, privacy, speed, cost and quality. It can let teams test systems, train models or share examples when real data cannot be widely exposed. That can reduce delays and support safer collaboration. But it can also create false confidence if people assume it matches reality well enough without checking. For a non-specialist, the key question is not 'Is it synthetic?' but 'Is it suitable for this purpose, and what risks remain?' That matters in procurement, governance, evaluation, cross-border data sharing, and any project where rare cases, bias or privacy are important.
How synthetic data is created and checked
Show that synthetic data comes from real data or rules, passes through a generation step, and must be evaluated before use.
01
1. Source material
Start with either real data, domain rules, or a simulation of a process. This source defines what patterns the generated records should reflect.
Draw three small input boxes labelled 'real data', 'rules', and 'simulation' feeding into one larger box.
02
2. Generation method
A statistical method or model creates new records that aim to preserve useful patterns without reproducing the original records exactly.
Place a central processing box labelled 'generator' with gears or transformation arrows inside.
03
3. Synthetic output
The result is a new dataset that may look like tables, text, images or sensor readings, depending on the task.
Show multiple output cards or rows emerging from the generator, labelled 'synthetic dataset'.
04
4. Evaluation
Check whether the output is useful for the task and whether privacy, bias and rare-case coverage are acceptable. If not, revise the source or method.
Add a final decision box with two badges: 'utility check' and 'privacy/risk check', plus a loop-back arrow to the generator.
A left-to-right flow with four stages: source material, generator, synthetic output, and evaluation for privacy and usefulness. Read it as a process: inputs shape the generated data, then checks determine whether it is suitable for training, testing or sharing.
Hover or tap each stage to reveal one short example, such as synthetic patient records, simulated traffic flows or fictional transaction tables.
Describe this without the diagram
The diagram explains synthetic data as a four-stage flow. First, source material is gathered: either real data, written rules, or a simulation of a real process. Second, a generator uses that source to create new records. Third, the output is a synthetic dataset, such as a table or set of text examples. Fourth, the synthetic data is checked for usefulness and for risks such as privacy leakage, bias and missing rare cases. If the checks fail, the process loops back and the method is adjusted.
With reduced motion: A static left-to-right flowchart shows source material feeding a generator, the generator producing a synthetic dataset, and an evaluation stage checking utility and privacy, with a simple return arrow back to the generator.
Synthetic data contains values produced artificially by rules, simulations or models rather than recorded directly from people, devices or events. The aim is to create records that reflect useful properties of a target population or process, such as ranges, relationships or frequencies, without simply copying source records.
Inputs may include a real dataset, business rules, or a simulation of how something works. A generation method then creates new examples. The output can be tables, text, images or other formats. Good synthetic data represents the properties needed for the intended task closely enough to be useful.
People use it for several reasons. It can help when real data is scarce, expensive, sensitive, imbalanced or hard to share across organisations or borders. It can also help test systems under rare conditions, such as unusual weather events or uncommon medical cases, if those cases are modelled carefully.
But synthetic data is not automatically safe or accurate. If source data has bias, errors or gaps, the synthetic version can reproduce them. If the generation method copies too closely, privacy risks can remain. If it smooths the data too much, rare but important cases may disappear. That is why it should be checked for utility, privacy, fairness and fitness for purpose before use.
How “Practice meeting notes for staff training” maps onto the real thing
| In the story | In the AI system |
|---|---|
| real note sets used to learn structure and patterns | original data or domain rules used to guide generation |
| fictional note sets created for practice | synthetic records generated to resemble real data |
| common agenda items and action-point style | statistical patterns and relationships preserved from the source |
| staff training before using approved real material | testing or training systems with lower access and privacy risk |
A worked example
The training officer needs a realistic practice dataset for a workshop on coding open-ended needs-assessment answers, while keeping the original responses inside the liaison office's approved environment.
Starting from: The approved source responses, coding guide and rules describing relevant features such as theme frequency, response length, missing answers and category combinations.
- An authorised analyst profiles the source responses inside the approved environment, recording the patterns needed for the exercise rather than selecting examples for direct reuse.
- A statistical or machine learning generator produces fictional answers designed to preserve those patterns, including less common coding categories and realistic missing fields.
- The analyst compares the synthetic and source datasets using utility measures such as category coverage, theme co-occurrence and coding performance.
- Disclosure checks look for verbatim passages, unusually close matches and rare combinations that could reveal or closely resemble a source response. Unsafe records are removed or regenerated.
- The training officer uses the cleared synthetic set in the workshop. Any final assessment of the coding process against real answers remains within the controlled environment.
Result: Colleagues can practise applying the coding guide to varied, realistic material without circulating the original answers. The exercise remains useful only if the synthetic set preserves relevant patterns and passes disclosure-risk checks.
In a United Nations workplace
A training officer in a liaison office is preparing a coding exercise for colleagues who will classify open-ended answers from a needs assessment. The exercise is due before the next morning's workshop, but the source answers contain details that should not appear in training materials.
Synthetic responses can reproduce useful features of the source material, such as theme frequencies, answer lengths, missing fields and combinations of categories. They allow colleagues to test the coding guide and compare coding decisions without working directly from respondents' records.
Before you try thisNeeds-assessment answers may contain personal, protection-related or otherwise sensitive details. They must not be uploaded to a synthetic-data tool that has not been approved for such material. Generated records are not automatically safe: models can reproduce distinctive wording or rare combinations, so outputs require disclosure-risk testing and clearance before circulation.
Where you may hear it
- During a vendor demo to an ICT governance board
- In an inter-agency working group reviewing a data-sharing protocol
- At a procurement meeting assessing test-data requirements
- During clearance of a pilot plan involving restricted records
Avoid confusion
Avoid confusion
Often confused with
Synthetic media
Synthetic media is text, images, audio or video generated or substantially altered by software, often AI, so it can resemble authentic material.
Data
Recorded observations, measurements or examples that people or systems can store, examine and use for a purpose.
Missing data
Some values that should be present in a dataset are absent, leaving blanks that can distort analysis and model behaviour.
Annotation and data labelling
People or software tools add task-relevant tags, notes or boundaries to data for model training, testing, analysis or monitoring.
Common misconceptions
Common misconception: Synthetic data is always anonymous and safe to share.
Correction: Not necessarily. If the generator memorises or reproduces original records too closely, sensitive information can still leak. Safety depends on how the data was generated, what tests were done, and what residual risk remains.
Common misconception: If it looks realistic, it will work for any AI task.
Correction: Visual or surface realism is not enough. The synthetic data must preserve the specific patterns needed for the task. A dataset can look convincing but still fail to represent rare cases, correlations or errors that matter in real use.
Common misconception: Synthetic data removes bias from the original data.
Correction: Usually it does not remove bias by itself. If the source data under-represents some groups or situations, the generated data may repeat that pattern or even amplify it unless the process is intentionally designed and evaluated to address it.
Quick check
Which statement best describes a sensible use of synthetic data?
Synthetic data can support development, testing and limited sharing, but it still needs evaluation. It is not automatically privacy-safe, it does not remove the need for appropriate validation, and it is not the same thing as anonymised original data.
Go deeper
Go deeper
The simple version
Synthetic data is artificially created data that resembles real data closely enough for practice, testing or some kinds of training.
The technical version
Synthetic data is artificially generated data produced from statistical or machine learning models, rules or simulations to approximate properties of a population or process. It may aim to preserve distributions, correlations, temporal patterns or structural constraints while reducing direct exposure of source records. Quality is judged by utility measures, such as downstream model performance, and risk measures, such as disclosure risk. It differs from anonymisation, which transforms existing records rather than creating new examples. Synthetic data and synthetic media can overlap. Generated images, audio, video or text are synthetic media, and they are synthetic data when used as data. Fidelity, privacy and rare-event coverage must still be evaluated.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This link gives a risk-management framework for evaluating AI systems and suits readers responsible for governance, assurance or procurement.
Intermediatelink checked 2026-07-31
Primary sourceOECD AI Principles (opens in a new tab)OECD
This link gives high-level principles for trustworthy AI and suits policy and programme readers who need a governance lens.
Intermediatelink checked 2026-07-31
EncyclopaediaSynthetic data (opens in a new tab)Wikipedia
This link gives a broad introductory overview of the term and suits readers who want a quick orientation before going deeper.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Synthetic data (opens in a new tab)YouTube search
This link is a YouTube search results page, so it helps readers browse different video explanations and suits those who learn well from short visual introductions.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Synthetic data (opens in a new tab)Reddit search
This link is a Reddit search results page, so it shows community questions and debates and suits readers who want to see common points of confusion and practical concerns.
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
Synthetic data can support testing, training and sharing, but it is not automatically anonymous. Its usefulness and disclosure risk must both be measured for the intended task.