03 · What You Need to Know
How Synthetic Data Are Created and Used in Research
What does the term synthetic data mean?
Synthetic data are artificially created information that can be used in place of, alongside, or separately from empirical observations for particular purposes.
The term covers a broad range of materials, including numerical tables, categorical records, text, images, and simulated events.
Some synthetic datasets are generated from models fitted to real data. Others are produced from mathematical assumptions or hypothetical scenarios.
For example, a researcher may create artificial student questionnaire records to test a statistical analysis script. Another may simulate sensor measurements under specified physical conditions.
Both datasets are synthetic because the individual records were generated rather than directly observed through the corresponding research activity.
The distinction between synthetic data and real research observations concerns their provenance rather than whether their values appear plausible.
How are synthetic data different from simulated data?
The terms overlap and are sometimes used interchangeably, but they can emphasize different aspects of artificial data generation.
Synthetic data commonly refers to artificially generated records intended to reproduce selected characteristics of a dataset or provide usable substitutes for particular tasks.
Simulated data often refers to observations generated from an explicitly specified model of a process, system, or hypothetical scenario.
For example, a model that generates artificial survey records resembling a reference dataset may be described as producing synthetic data.
A computational model that produces hypothetical measurements under specified physical equations may be described as a simulation.
However, these categories are not mutually exclusive. Simulated observations are often synthetic, and synthetic-data generation may rely on simulation.
The more useful distinction concerns the generation assumptions, intended purpose, and claims supported by the output.
The comparison between synthetic and simulated data should therefore be interpreted as a methodological distinction rather than an absolute division.
How can AI generate synthetic data?
AI systems can generate artificial data through several computational approaches.
Some models learn statistical patterns from a reference dataset and produce new records intended to resemble its characteristics.
Other systems generate structured information from prompts describing hypothetical observations.
Generative models may also be conditioned on particular variables, scenarios, or constraints.
| Generation approach |
How it works |
Typical consideration |
| Statistical simulation |
Samples observations from specified probability distributions or mathematical models. |
Validity depends on the assumptions and implementation. |
| Generative adversarial networks |
Train competing neural networks to learn characteristics of reference data. |
May struggle with training instability or incomplete representation of rare cases. |
| Variational autoencoders |
Learn a probabilistic latent representation from which artificial records can be generated. |
May smooth or inadequately reproduce some data characteristics. |
| Diffusion-based models |
Learn generative processes that progressively construct samples from noise. |
Performance depends on the data modality, model design, and training conditions. |
| Large language models |
Generate textual or structured records conditioned on instructions and context. |
Plausible outputs do not automatically satisfy statistical or empirical validity requirements. |
These methods are not interchangeable. Their suitability depends on the type of data, the available source information, computational requirements, and intended research use.
How do generative adversarial networks produce synthetic data?
Generative adversarial networks, commonly called GANs, were introduced by Goodfellow and colleagues in 2014.
A conventional GAN consists of a generator and a discriminator trained through an adversarial process.
The generator produces artificial samples, while the discriminator attempts to distinguish generated samples from reference data.
Through training, the generator may learn to produce outputs resembling selected characteristics of the training distribution.
GANs have been investigated for image generation and other forms of synthetic-data production.
However, successful training does not guarantee that every relevant property of the original dataset is preserved.
Generated samples may inadequately represent rare cases, exhibit limited diversity, or retain information from training examples.
Researchers should validate the resulting data against the requirements of their intended application.
How do variational autoencoders generate synthetic records?
Variational autoencoders, or VAEs, are generative models that learn a probabilistic representation of data through latent variables.
During training, the model learns relationships between observed data and a lower-dimensional latent representation.
New artificial observations can then be generated by sampling from the learned latent distribution and decoding those samples.
VAEs may be useful for generating continuous or structured data, depending on the architecture and training procedure.
However, generated records may not preserve every important characteristic of the reference data.
For example, some models may inadequately reproduce extreme values, multimodal distributions, or complex dependencies.
Validation should therefore consider the characteristics that matter for the intended research task rather than relying solely on general similarity.
How do diffusion models generate synthetic data?
Diffusion models learn a generative process associated with progressively adding noise to data and learning how to reverse that process.
During generation, a model can produce new samples through a sequence of denoising steps.
Diffusion-based methods have become prominent in image generation and have also been investigated for other data modalities.
Depending on the architecture, they may generate synthetic images, signals, or structured records.
However, the quality and scientific usefulness of generated samples depend on the training data, model assumptions, and evaluation methods.
A visually realistic synthetic scientific image does not establish that the depicted observation occurred experimentally.
Researchers should distinguish generated imagery used for methodological testing from images presented as genuine scientific evidence.
How can large language models generate synthetic datasets?
Large language models can produce artificial textual and structured records when given instructions describing variables, scenarios, or hypothetical respondents.
For example, a researcher might ask a model to generate sample questionnaire responses using specified response categories.
The resulting records may be useful for demonstrating data formats or testing software.
However, general-purpose language models do not necessarily implement a statistically validated sampling procedure when generating numerical tables.
Their outputs may contain inconsistencies, unintended patterns, or relationships shaped by prompt wording.
For applications requiring controlled statistical characteristics, verified simulation code or purpose-built synthetic-data models may be more appropriate.
AI can assist in developing those procedures, but researchers should independently verify the resulting implementation.
Do synthetic data always require real training data?
No. Some synthetic-data methods learn from existing empirical datasets, while others generate observations from theoretical assumptions or specified rules.
For example, a simulation of a physical process may produce artificial measurements using mathematical equations without training a generative model on a participant dataset.
Similarly, a software-testing dataset may be constructed from predefined variable types and allowable values.
Data-driven generative models, by contrast, often rely on reference data to learn distributions and relationships.
The distinction matters because synthetic datasets derived from sensitive records may introduce privacy risks that do not arise in the same way when data are generated entirely from nonconfidential assumptions.
Researchers should document whether empirical source data were used and what information the generation process depended upon.
Can synthetic data reproduce the statistical properties of real datasets?
Some methods can reproduce selected characteristics of reference data, including marginal distributions, correlations, or other relationships.
However, similarity is rarely complete across every relevant dimension.
A synthetic dataset may approximate overall averages while inadequately representing rare cases or minority subgroups.
It may preserve individual variable distributions while failing to reproduce complex relationships among variables.
These limitations matter when the data are used for prediction, subgroup analysis, or other tasks sensitive to the properties that were not preserved.
Researchers should evaluate whether the generated dataset retains the characteristics necessary for its intended application.
The question of validating synthetic data therefore requires task-specific evaluation rather than a single universal similarity score.
What are the main research applications of synthetic data?
Synthetic data may support several legitimate research activities.
They can provide controlled inputs for testing software, evaluating statistical procedures, developing computational models, or demonstrating analytical methods.
They may also help researchers investigate hypothetical scenarios that are difficult or impractical to observe directly.
In some circumstances, synthetic datasets can reduce the need to share original participant-level records.
However, the appropriate use depends on the research question and the properties of the generated material.
A dataset suitable for testing a data import function may not be suitable for estimating population prevalence.
Likewise, synthetic participant responses may help test a coding procedure without providing genuine evidence of human attitudes or experiences.
The legitimacy of synthetic data in research depends on the methodological purpose and the claims made from the data.
Can synthetic data be used to test research code?
Yes. Software testing is one of the more straightforward applications of synthetic data.
Researchers can create artificial records with known properties to verify that data-cleaning scripts, statistical functions, and visualization procedures behave as intended.
For example, a synthetic survey dataset may include reverse-coded questionnaire items, missing values, and categorical variables to test an analysis pipeline.
The objective is to verify the software rather than draw empirical conclusions about the fictional respondents.
Researchers should nevertheless ensure that the test data exercise the relevant conditions and that expected outputs are independently specified where possible.
The use of synthetic data to test research analysis code can support reproducibility when test cases and expected results are preserved.
Can synthetic data support prediction model development?
Potentially. Synthetic data may help researchers develop or evaluate prediction models when genuine data are limited, imbalanced, or difficult to access.
However, model performance on artificial observations does not automatically establish performance on real-world data.
Generated records may reproduce biases, omit rare cases, or fail to represent important relationships in the intended deployment population.
Researchers should evaluate the resulting prediction model using appropriate independent reference data whenever the intended claim concerns real-world performance.
They should also consider whether the synthetic-data generator and prediction model share information that could create misleading evaluation results.
The question of synthetic data for prediction model development therefore involves both data quality and evaluation design.
Can synthetic data protect privacy?
Synthetic data may reduce direct exposure of original participant-level records, but artificial generation does not automatically guarantee anonymity.
A generative model trained on sensitive information may reproduce distinctive records or reveal information about individuals represented in its training data.
Privacy risk depends on the generation method, source data, model behavior, and information available to potential recipients.
Researchers should therefore assess disclosure risks before releasing synthetic datasets derived from confidential information.
Techniques such as differential privacy may provide formal privacy guarantees when correctly implemented, but those guarantees depend on the mechanism and its parameters.
The separate question of whether synthetic data protect research participant privacy should not be answered solely by noting that the released records are artificial.
How should researchers evaluate synthetic-data quality?
Synthetic-data quality should be assessed in relation to the intended application.
For data intended to resemble a reference dataset, relevant evaluations may include distributional similarity, relationships among variables, subgroup representation, and downstream analytical performance.
For simulation studies, researchers should verify that the generated observations follow the specified mathematical or computational assumptions.
For software testing, useful evaluation may concern whether the dataset includes the intended edge cases and produces expected program behavior.
Where sensitive source data are involved, privacy risk should be evaluated separately from statistical similarity.
A dataset can preserve useful statistical patterns while still posing disclosure risks.
Likewise, a privacy-preserving dataset may have limited utility for certain analyses.
Researchers should recognize these tradeoffs rather than assume that a single quality measure captures every relevant property.
Are AI-generated synthetic data the same as AI-fabricated research data?
No. Artificial generation and research fabrication are not synonymous.
The United States Office of Research Integrity defines fabrication as making up data or results and recording or reporting them.
The research-integrity problem arises when invented observations are falsely represented as genuine evidence.
Synthetic data may be entirely legitimate when their artificial origin is clear and their use is scientifically appropriate.
For example, a generated dataset used to demonstrate a statistical procedure is different from a generated dataset falsely reported as responses collected from actual participants.
The distinction between synthetic and fabricated data depends on how the records are represented and used.
How should synthetic data be documented and reported?
Researchers should explain how synthetic data were generated and what purpose they served.
Relevant documentation may include the generation method, source data where applicable, software or model version, parameters, assumptions, validation procedures, and limitations.
When artificial and empirical records coexist, their provenance should remain distinguishable throughout analysis.
Researchers should also report whether the synthetic dataset was used for software testing, simulation, model development, or another methodological activity.
Where AI systems were involved, applicable journal and institutional disclosure requirements should be followed.
The separate issue of reporting synthetic data in research papers concerns providing enough detail for readers to understand and evaluate the generation procedure.
