{
  "id": "synthetic-data",
  "title": "Synthetic Data",
  "lang": "en",
  "source_lang": "en",
  "summary": "Synthetic data are artificially generated observations used to train models, test systems, simulate conditions, and support analysis while controlling selected properties of the data.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "concept",
  "aliases": [
    "Artificial data",
    "Synthetically generated data"
  ],
  "generated_at": "2026-11-06",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/synthetic-data/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/synthetic-data/",
      "json": "https://aiwiki.page/zh/synthetic-data.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Synthetic data are observations, records, or other information produced by a computational process rather than obtained solely through direct measurement of the events or entities they represent. They may resemble existing datasets, describe hypothetical populations, or depict situations that have never occurred. Examples include simulated road scenes, artificial financial transactions, generated medical records, and machine-written instruction–response pairs. Synthetic datasets can be created from explicit rules, statistical models, physical simulations, or generative models learned from real observations.\n\nSynthetic data are important in machine learning, software testing, scientific research, and statistical data dissemination. They can supply labels that would otherwise be expensive to obtain, expose systems to rare conditions, and make restricted information more accessible. Their value depends on the intended use: a dataset suitable for testing a database interface may be unsuitable for estimating population health, while realistic-looking images may fail to improve a vision system.\n\nArtificial generation does not by itself establish accuracy, anonymity, representativeness, or legal permission to use the output. Synthetic data inherit assumptions from their construction process and may preserve defects or sensitive information from their source material. Their study therefore combines generation methods with evaluation, privacy analysis, provenance, and controls on downstream use."
    },
    {
      "heading": "Definition, scope, and principal distinctions",
      "text": "The defining feature of synthetic data is their method of production, not their file format or apparent realism. A synthetic observation is constructed by a model or procedure to stand for an observation that might otherwise have been collected. The underlying object may be a row in a table, an image, a waveform, a document, a sequence of transactions, or an entire interacting population. Some generators imitate a particular source dataset; others implement a theory or specification without learning from any individual-level records.\n\nA useful distinction separates fully synthetic from partially synthetic data. In statistical disclosure work, a fully synthetic release generally replaces the sensitive observed values intended for release with generated values, sometimes using a sampling frame that is itself not confidential. A partially synthetic release retains some observed values and synthesizes selected fields or records. Terminology varies between disciplines: in commercial systems, “fully synthetic” may mean simply that every output row is generated, even though the generator was trained on confidential records. Consequently, the label alone says little about privacy protection.\n\nSynthetic data also differ from anonymized, masked, and pseudonymized data. Replacing a person’s name with an arbitrary identifier transforms an existing record rather than necessarily generating a new observation. Adding noise to a numerical field may constitute perturbation, synthesis, or both, depending on the procedure. A generated table can contain no exact copies yet still reveal source information through unusual combinations of attributes. Conversely, a useful test dataset can deliberately contain recognizable fictitious names because it is designed for functionality rather than statistical resemblance.\n\nData augmentation overlaps with synthesis but is not identical to it. Rotating a photograph or adjusting its brightness modifies an existing example while usually preserving its label. Generating a new scene with different objects constructs a broader set of properties. The boundary is gradual: combining image regions, interpolating feature vectors, or generating paraphrases can reasonably be described as either augmentation or synthesis. What matters operationally is which information is inherited, which is newly created, and which relationships the transformation preserves.\n\nSynthetic datasets may target resemblance to an observed probability distribution, coverage of specified scenarios, or satisfaction of logical constraints. These objectives need not coincide. A statistically representative transaction dataset might contain almost no major fraud cases, whereas a stress-test dataset may deliberately overrepresent them. Quality must therefore be defined relative to a declared purpose rather than treated as a single, universal measure of realism."
    },
    {
      "heading": "Historical development",
      "text": "Synthetic data have several historical lineages. Scientific simulation long preceded contemporary generative models: researchers produced artificial observations to understand physical systems, study measurement error, and evaluate statistical procedures. The Monte Carlo method, developed prominently during the 1940s, uses repeated random sampling to estimate quantities that may be difficult to calculate directly. Simulated samples made it possible to compare estimators under known conditions, including situations where the correct population parameters were available by construction.\n\nAnother lineage arose from confidentiality in official statistics. Public-use microdata create a tension between analytical detail and disclosure risk. Detailed records support research, but combinations such as age, location, occupation, and household composition can identify individuals even when names are absent. Donald Rubin’s 1993 proposal for synthetic public-use microdata helped establish an approach in which agencies would release generated records rather than simply modifying confidential observations. Jerome Reiter’s work in the early 2000s further developed inference and disclosure assessment for partially synthetic releases.\n\nThese developments were closely connected to methods for handling missing data. In both settings, a model produces plausible values conditional on observed information. The purpose differs, however: missing-data methods address incomplete observations, whereas disclosure-oriented synthesis deliberately replaces information to reduce the connection between released records and identifiable respondents. Multiple generated datasets can express uncertainty associated with synthesis, although the appropriate inferential rules depend on the construction and release design.\n\nIn machine learning, artificial examples became a practical response to small datasets and class imbalance. The Synthetic Minority Over-sampling Technique, or SMOTE, introduced by Nitesh Chawla and colleagues in 2002, creates minority-class examples through interpolation between nearby observations. Rather than duplicating a rare example, it constructs additional feature vectors along local line segments. This can help some classifiers, but it assumes that the interpolated region belongs to the same class and is meaningful in the chosen representation.\n\nThe development of deep learning expanded the range of data that could be synthesized. Generative adversarial networks, introduced in 2014, made learned image generation a major research direction. Variational autoencoders, autoregressive models, and later diffusion models offered other approaches. During the 2020s, capable large language models greatly increased the practical production of synthetic text, code, and instructions. These developments broadened access to generation while making source attribution, evaluation contamination, and repeated learning from generated material more consequential."
    },
    {
      "heading": "Statistical and learned generation methods",
      "text": "A statistical generator specifies a process for drawing artificial observations. For a table with variables \\(X_1,\\ldots,X_d\\), one approach factors their joint distribution into conditional distributions:\n\n\\[p(x_1,\\ldots,x_d) = \\prod_{j=1}^{d}p(x_j\\mid x_1,\\ldots,x_{j-1}).\\]\n\nEach variable is generated in sequence, conditional on previously generated values. The conditional models might be regressions, decision trees, or more flexible estimators. The ordering and modeling assumptions affect which dependencies are reproduced. Independently sampling each column may preserve its marginal distribution while destroying relationships such as those between household income, household size, and housing expenditure.\n\nRule-based generators instead encode explicit requirements. A software-testing system might generate account balances within prescribed ranges, ensure that transaction dates follow account-opening dates, and assign valid references between tables. Rules can enforce properties that a statistical model might violate, but manually authored dependencies are difficult to maintain in complex domains. Hybrid systems therefore combine learned distributions with constraints, rejection sampling, or repair procedures.\n\nNeural generators represent distributions through trainable functions. A variational autoencoder maps examples into a probabilistic latent space and learns a decoder that reconstructs or generates observations. A generative adversarial network trains a generator against a discriminator that attempts to distinguish generated from real examples. Diffusion models learn to reverse a sequence of noise-corruption steps. Autoregressive models generate components sequentially, conditioning each on preceding components. These approaches differ in training stability, sampling cost, controllability, and their ability to represent discrete or continuous structures.\n\nConditional generation supplies additional information, such as a class label, a text description, or a required scenario. If \\(z\\) is a random input and \\(c\\) is a condition, generation can be expressed as\n\n\\[\\tilde{x}=G_{\\theta}(z,c),\\]\n\nwhere \\(G_{\\theta}\\) is a generator with parameters \\(\\theta\\). The condition can direct an image generator toward a rainy road scene or a table generator toward a particular age group. Conditioning does not guarantee compliance: a generator may ignore difficult conditions, satisfy only superficial features, or produce inconsistent combinations.\n\nRelational, temporal, and network data introduce constraints beyond individual records. A plausible customer table is insufficient if its purchase table violates foreign-key relationships or produces impossible event sequences. Time-series synthesis must address seasonality, temporal dependence, abrupt regime changes, and irregular sampling. Network synthesis may need to preserve degree patterns, community structure, and interactions over time. These cases illustrate why a generator that works well for independent rows may be unsuitable for an organization’s complete database."
    },
    {
      "heading": "Simulation, rendered environments, and exact labels",
      "text": "Simulation produces synthetic data by modeling the process that would generate observations. In computer vision, a virtual scene specifies geometry, materials, lighting, objects, and sensor parameters. Computer graphics then renders images from selected viewpoints. Because the scene is constructed explicitly, the system can also output depth, object identities, surface normals, and pixel-level class labels without manual annotation. A single scene can therefore support object detection, semantic segmentation, and motion estimation.\n\nThe SYNTHIA and Virtual KITTI datasets, both introduced in 2016, are examples of synthetic environments used for autonomous-driving research. Such datasets allow controlled changes to weather, illumination, traffic, and viewpoint. Rather than waiting to collect a particular combination on public roads, a developer can render it repeatedly. The same underlying scene can also be observed by several simulated sensors, enabling aligned images, depth measurements, and other signals.\n\nThe principal difficulty is the difference between simulated observations and the deployment environment. Rendering may omit lens flare, rolling-shutter distortion, sensor noise, irregular materials, or behaviors that occur in real traffic. This is a form of distribution shift often discussed as the simulation-to-reality gap. A system can perform well on generated images yet fail when the real sensor produces patterns absent from the simulator.\n\nDomain randomization addresses this problem by varying simulation parameters broadly. Object textures, camera positions, lighting, friction, or physical dimensions may be sampled across ranges wider than those expected in deployment. The aim is not always to make every image photorealistic; it is to prevent the model from depending on narrow, artificial regularities. Domain adaptation offers complementary methods for aligning simulated and real representations. Neither approach eliminates the need for evaluation using real observations.\n\nSimulation is also central to robotics and reinforcement learning. An agent can generate trajectories by acting in an environment, receiving observations and rewards. This makes training possible when real-world experimentation would be slow, costly, or hazardous. Scientific simulators similarly produce synthetic sensor readings, spectra, and physical fields. Here the quality of the data depends on model equations, numerical approximations, boundary conditions, and parameter estimates. Exact labels are exact only with respect to the simulator: a perfectly annotated collision in a virtual environment does not establish that the modeled dynamics match a physical collision."
    },
    {
      "heading": "Synthetic text, code, and instruction datasets",
      "text": "Text synthesis can produce individual sentences, documents, conversations, structured answers, or executable programs. Earlier approaches often relied on templates and grammars. Contemporary systems frequently use a pretrained language model to generate examples conditioned on instructions, source documents, or desired task categories. This makes it possible to create large collections for natural language processing, but the generator’s fluency can conceal factual errors, repetitive reasoning, and unsupported assumptions.\n\nOne prominent use is instruction tuning: generated tasks and responses become training data for a model expected to follow instructions. The Self-Instruct method, described in 2022, expanded a small collection of human-written tasks through model-generated instructions and instances, followed by filtering. The resulting examples were used for fine-tuning. The approach demonstrated how synthetic generation could broaden task coverage without requiring every example to be authored individually, while leaving quality dependent on the seed tasks, generation procedure, and filtering rules.\n\nSynthetic code offers stronger opportunities for automated checking than unrestricted prose. A generated program can be compiled, executed, and tested against specified input–output behavior. Mathematical answers can sometimes be checked computationally, and formal statements may be verified with a proof assistant. However, successful execution is not equivalent to full correctness. A program may pass limited tests while failing on untested cases, and a verifier establishes only the properties represented in its specification.\n\nTeacher–student pipelines use one model to generate data for another. This overlaps with knowledge distillation, although generated text targets differ from transferring a teacher’s full probability outputs. Synthetic explanations, alternative answers, and preference comparisons can provide training signals. If the teacher repeatedly makes the same error, a larger generated corpus can amplify that error rather than correct it. Multiple generators, diverse prompts, and external verification can change the error structure, but their effectiveness must be measured rather than presumed.\n\nDataset construction also requires attention to duplication and evaluation boundaries. A generator may reproduce known benchmark questions or create close paraphrases of them. If these enter training, later performance can reflect benchmark data contamination rather than improved ability. Long synthetic conversations may contain contradictions across turns; generated citations may refer to nonexistent sources; and an answer accepted by another model may merely share that model’s preferences. Provenance and task-specific verification are therefore more informative than the sheer number of generated tokens."
    },
    {
      "heading": "Applications across research and industry",
      "text": "In software development, synthetic data support functional testing, load testing, and reproducible demonstrations. A payment service can be tested with valid and invalid transaction sequences without exposing actual customers. A relational database can be populated with millions of artificial rows to investigate query performance, indexing, and storage behavior. For these purposes, realistic record lengths, skew, relationships, and transaction patterns may matter more than a visually convincing sample of rows.\n\nMachine-learning applications focus on training coverage and labels. Synthetic examples can increase the representation of rare classes, create controlled combinations of attributes, or supply annotations unavailable from ordinary collection. For instance, a rendered scene can associate every image pixel with a known object. Yet increased dataset size is not equivalent to increased independent information. Repeated variations of a small set of source patterns may produce limited gains in generalization.\n\nSensitive-data applications include sharing artificial versions of financial, educational, and health records. Such releases can support exploratory analysis, method development, or preparation of code before access to restricted data. In some statistical arrangements, researchers use a synthetic file to develop an analysis and then submit the code for execution against confidential records. This separates development convenience from the evidential status of final results. A synthetic release need not reproduce every association well enough to support direct substantive inference.\n\nIn cybersecurity, artificial event logs and network traffic can exercise detection systems. A generator may simulate login sequences, malicious requests, or unusual access patterns. The difficulty is reproducing the background behavior against which an attack must be detected. A classifier can appear successful when synthetic attacks contain formatting artifacts or unrealistic timing, yet fail against actual adversaries. Similar problems arise in fraud detection when generated fraudulent transactions are easier to distinguish than real ones.\n\nScientific applications use synthetic observations to test analysis pipelines under controlled assumptions. Astronomers inject artificial sources into images to estimate detection completeness; researchers simulate sequencing reads to assess analytical tools; and imaging studies use modeled measurements to investigate reconstruction methods. These uses differ from creating new empirical evidence about the world. Synthetic observations show what an instrument or method would produce under specified conditions. Their evidential value depends on how well those conditions and the generating model correspond to the phenomenon being studied."
    },
    {
      "heading": "Evaluating fidelity, utility, and coverage",
      "text": "Evaluation begins with the intended task. Statistical fidelity asks whether generated data reproduce relevant properties of a reference population. Utility asks whether they support a particular analysis or system. Coverage asks which conditions, groups, and failure cases they include. These dimensions can conflict: a dataset that overrepresents rare incidents may be useful for testing but unsuitable for directly estimating their frequency.\n\nFor tabular data, basic checks include distributions of individual variables, missingness patterns, correlations, conditional relationships, and constraint violations. Matching all one-dimensional histograms is insufficient because datasets with identical marginals can have very different joint structures. Checks may therefore examine relationships such as expenditure conditional on income or event rates within age groups. In relational datasets, evaluation must also consider aggregation across tables and the number and ordering of events associated with each entity.\n\nA common utility test is “train on synthetic, test on real.” A model trained using generated examples is evaluated on a real test set that was not used to fit the generator, choose its settings, or filter its outputs. Performance can be compared with a real-data training baseline and with mixed real–synthetic training. Different downstream models may produce different conclusions, so a single classifier is not a universal measure of synthetic quality. Proper separation is necessary to avoid data leakage.\n\nDistributional distances and discrimination tests provide additional evidence. A classifier trained to distinguish real from synthetic samples can reveal differences, but its success depends on model capacity, sample size, and representation. Failure to distinguish the samples does not prove that their distributions are identical. Image-generation metrics based on pretrained feature extractors similarly measure selected properties of those representations rather than all aspects of visual or semantic quality.\n\nUncertainty and subgroup behavior require separate examination. If synthetic datasets support statistical inference, analysts must account for the uncertainty introduced by fitting and sampling from the generator. Treating generated records as independent new measurements can create unjustifiably narrow confidence intervals. Performance should also be examined for small groups and distributional tails, where a generator may smooth away important variation. A dataset can achieve good average utility while substantially misrepresenting a rare category, an extreme value, or a consequential interaction between variables."
    },
    {
      "heading": "Privacy protection and disclosure risk",
      "text": "Synthetic data can reduce direct exposure of original records, but artificial generation is not a privacy guarantee. A generator may memorize source examples, reproduce rare combinations, or reveal whether a person participated in training. Privacy analysis therefore depends on a threat model: what information an adversary already has, what they can access, and what they are trying to learn. A public dataset, a restricted research environment, and an interactive generation service create different opportunities for attack.\n\nA membership inference attack attempts to determine whether a particular record was included in training. Other attacks seek sensitive attributes from known characteristics or attempt to reconstruct information about source records. Removing exact duplicates addresses only one class of exposure. An unusual combination of occupation, location, and household characteristics may remain recognizable even when every generated row differs from every original row.\n\nDistance-based privacy checks are useful but limited. Measuring the nearest synthetic neighbor of each real record can identify suspiciously close reproductions. However, closeness may be normal in a dense population, and distant records may still reveal a sensitive relationship. Comparing distances to both training and held-out real data can improve interpretation, but no single empirical distance measure establishes a general privacy guarantee.\n\nDifferential privacy provides a formal framework. A randomized mechanism \\(M\\) is \\((\\varepsilon,\\delta)\\)-differentially private if, for neighboring datasets \\(D\\) and \\(D'\\) differing in one protected contribution and every measurable output set \\(S\\),\n\n\\[\\Pr[M(D)\\in S] \\leq e^{\\varepsilon}\\Pr[M(D')\\in S]+\\delta.\\]\n\nThe guarantee limits how much the output distribution changes because of that contribution. The neighboring-dataset definition matters: protecting one row is not necessarily equivalent to protecting a person who contributes many rows.\n\nA generator can be fitted with a differentially private mechanism and then sampled without repeatedly accessing confidential data. Under differential privacy’s post-processing property, those samples do not consume additional privacy budget merely because more are produced. However, private-data-dependent preprocessing, selection, validation, or repeated fitting must also be included in the analysis. Stronger protection can reduce fidelity, particularly for small groups. Differential privacy also does not prevent every population-level inference or establish that the data are suitable for every legal or ethical purpose."
    },
    {
      "heading": "Production pipelines, provenance, and governance",
      "text": "A production pipeline starts with a specification rather than a generator. The specification identifies the unit of observation, target population, allowed uses, required relationships, and acceptable errors. A dataset for interface testing may require valid formats and referential integrity; one for statistical research may require preservation of multivariate relationships. These requirements determine what source information is needed and how success can be assessed.\n\nSource preparation includes schema inspection, treatment of missing values, handling of outliers, and separation of training and evaluation data. These operations can substantially alter the resulting dataset. Removing rare records may reduce some disclosure risks while also erasing legitimate cases; standardizing categories can conceal inconsistencies that matter in deployment. Preparation decisions are therefore part of the synthesis model, not merely administrative cleanup.\n\nGeneration is commonly followed by validation and filtering. Checks can reject impossible dates, incompatible units, invalid categories, or broken relationships. Generated code may be executed in a sandbox, while generated text may be checked against source documents or task rules. Filtering changes the distribution of accepted outputs. If a validator disproportionately rejects examples from a particular group or scenario, the final dataset can be biased even when the initial generation process was balanced.\n\nData lineage records how source datasets, generator versions, prompts, simulation settings, random seeds, and filters contributed to a release. Metadata can identify which fields are synthetic, which remain observed, and which labels come directly from a simulator. Reproducibility may require more than a random seed because hardware, library versions, nondeterministic operations, and externally hosted models can change outputs. Versioned generation configurations and recorded evaluation results make later comparisons more interpretable.\n\nGovernance also concerns access, retention, licensing, and documentation. Synthetic output does not automatically remove obligations associated with confidential source material or intellectual property. The legal status depends on the data, the jurisdiction, the generation process, and the proposed use. Operational controls may still include access control, release review, and restrictions on redistribution. Documentation can distinguish a dataset approved for software testing from one evaluated for a particular analytical task, preventing a narrow quality claim from being mistaken for general equivalence to real data."
    },
    {
      "heading": "Fundamental limits and feedback effects",
      "text": "A generator cannot recover information absent from its sources or assumptions merely by producing more samples. If a training dataset excludes a population or a relevant failure mode, learned synthesis will generally not reconstruct it reliably without additional information. Conditional prompts and simulation rules can introduce new scenarios, but their credibility comes from those added assumptions, domain knowledge, or external evidence—not from sample count alone.\n\nThe distinction is important for causation. A generator trained on observational associations may reproduce \\(p(y\\mid x)\\) without representing what would happen if an intervention changed \\(x\\). For example, an association between treatment and poor outcomes can reflect the fact that sicker patients receive treatment. Generating more records with that association does not establish the treatment’s causal effect. Intervention studies based on synthetic data require a defensible causal model, not only accurate observational resemblance.\n\nRepeated learning from generated outputs creates another limitation. Model collapse refers to degradation that can occur when successive models learn from earlier models’ generated data, particularly when generated samples replace original observations and rare events are progressively lost. Theoretical and empirical work in 2023 and 2024 examined this behavior. It is not an inevitable consequence of using any synthetic data: outcomes depend on the retention of real data, sampling procedures, model capacity, filtering, and the mixture used during training.\n\nFeedback can also arise without repeated model generations. If a generator and evaluator share similar training sources or architectures, the evaluator may reward familiar artifacts rather than genuine correctness. Selecting only high-scoring samples can narrow diversity and encourage reward hacking. A model trained on the selected dataset may then perform well under the same evaluator while becoming less robust outside that evaluation loop.\n\nRare events and open-ended environments remain particularly difficult. A simulator can enumerate known hazards but cannot ensure coverage of unanticipated ones; a generative model can produce diverse-looking examples while missing important modes of the underlying distribution. Synthetic evidence is therefore strongest when its assumptions, coverage, and validation targets are explicit. For tasks involving real-world decisions, a central unresolved problem is determining how much confidence performance on generated conditions warrants when the actual deployment distribution is only partially known."
    }
  ],
  "links": [
    "machine-learning",
    "data-augmentation",
    "probability-distribution",
    "monte-carlo-method",
    "statistics",
    "deep-learning",
    "generative-adversarial-network",
    "variational-autoencoder",
    "diffusion-model",
    "large-language-model",
    "decision-tree-learning",
    "latent-space",
    "foreign-key",
    "database",
    "computer-vision",
    "computer-graphics",
    "object-detection",
    "semantic-segmentation",
    "distribution-shift",
    "domain-adaptation",
    "reinforcement-learning",
    "language-model",
    "natural-language-processing",
    "training-data",
    "fine-tuning",
    "proof-assistant",
    "knowledge-distillation",
    "benchmark-data-contamination",
    "relational-database",
    "generalization",
    "cybersecurity",
    "test-set",
    "data-leakage",
    "confidence-interval",
    "membership-inference-attack",
    "differential-privacy",
    "sandbox-computer-security",
    "data-lineage",
    "metadata",
    "intellectual-property",
    "access-control",
    "causation",
    "model-collapse",
    "reward-hacking"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}