What synthetic data is
Synthetic data is artificially generated datasets that reproduce the statistical properties and patterns of real data, without containing information about specific people or entities. It isn’t randomly invented data: it’s produced from models trained on original data, with the aim of generating a set that is statistically representative and functional for a specific purpose.
As the Spanish Data Protection Agency (AEPD) notes in its document on the subject, for a dataset to be considered synthetic it has to preserve the characteristics and properties of the real data for that specific use case. If it doesn’t meet that minimum utility threshold, it can’t be classified as valid synthetic data.
In practice, this means a synthetic-data model can reproduce, for example, the age distribution of a patient cohort, the transaction patterns of a customer portfolio or the failure records of a production line, without any of that data corresponding to a real person or event. The value is that it preserves the internal logic of the original phenomenon.
Types of synthetic data: not all are the same
Before assessing whether synthetic data fits a project, it’s worth distinguishing between the three main types, because their uses and guarantees differ.
Fully synthetic data
These are datasets generated entirely artificially, with no real-world record. They’re built by estimating the statistical distributions of the original set and generating new samples from them.
Their main advantage is maximum privacy protection; their limitation is that they can lose variability in highly heterogeneous or infrequent phenomena. They’re the usual resort for software development and testing environments where access to real data is completely restricted.
Partially synthetic data
They start from real data but replace only the most sensitive attributes (names, identifiers, contact details, identifiable clinical information) with artificial values. The rest of the record keeps its original structure.
They’re especially useful when you need to preserve the coherence of the business data but can’t share the personal attributes, for example when collaborating with an external supplier or passing data to a third-party analytics environment.
Hybrid synthetic data
They combine real records with fully synthetic records. They’re used when the volume of real data is insufficient and you need to expand the training set while keeping part of the original variability. It’s the most common form in machine learning projects where real data exists but is scarce.
How synthetic data is generated
The generation method largely determines the quality and reliability of the result. There’s no single approach valid for every case.
- Classical statistical models analyze the distributions of the original set and generate samples from them. They’re suitable when the relationships between variables are known and well defined, and when the dataset isn’t excessively complex. Their main advantage is interpretability: it’s easier to audit why the model generates what it generates.
- Generative adversarial networks (GANs) operate with two neural networks in competition: one that generates data and another that tries to tell it apart from the real data. This iterative competition produces high-fidelity results, especially useful for image data or sets with very complex relationships between variables. They’re currently the most-used method in healthcare and computer vision projects, though they require large volumes of starting data and significant computational capacity.
- Variational autoencoders (VAEs) learn a compressed representation of the original data and generate new samples from it. They’re more stable than GANs during training and produce more controllable results, though their fidelity on highly heterogeneous data can be lower.
Choosing the method isn’t a minor technical decision. It depends on the type of data, the volume available, the level of privacy required and the final purpose of the synthetic set.
When it makes sense to use synthetic data
Synthetic data adds real value in three specific situations. Outside them, it’s worth evaluating other alternatives before assuming it’s the right solution.
- When the volume of real data is structurally low. In rare diseases, very low-frequency industrial events or fraud scenarios with few recorded cases, generating synthetic data makes it possible to expand the training set and improve the model’s robustness. It doesn’t remove the need for real data, but it reduces the penalty for scarcity.
- When regulation or privacy blocks access to the data. The General Data Protection Regulation (GDPR) and sector-specific rules in healthcare, finance or the pharmaceutical industry impose significant restrictions on what data can be shared between organizations, used in development environments or transferred to third parties. Synthetic data makes it possible to collaborate and develop without exposing personal or confidential information, acting as an effective privacy technique, as the AEPD itself acknowledges.
- In development, testing and validation environments. Before connecting an AI model to real data in production, it makes sense to test the architecture, the pipelines and the model logic with synthetic data. It reduces risk, speeds up development cycles and allows iteration without handling sensitive information in non-secured environments.
Sectors such as logistics, manufacturing and financial services have spent years incorporating synthetic data into their development cycles for predictive models, anomaly detection and demand planning, precisely for these reasons.
When synthetic data is not the solution
This is the point most articles on the subject develop least, and it’s the one with the greatest impact on decision-making.
- Synthetic data doesn’t fix quality problems at the source. If the starting real data is incomplete, biased or poorly structured, the generative model will learn exactly those same patterns and reproduce them at scale. A synthetic set built on deficient data doesn’t expand the problem: it multiplies it. Before considering generating synthetic data, it’s essential to fix the quality of the source data.
- They don’t replace real data when individual variability matters. There are projects where the model needs to learn from the real heterogeneity of a population (unusual clinical variations, atypical customer behavior, infrequent but critical machine failures), and synthetic data, however well generated, introduces a degree of homogenization that can penalize the model’s performance in production.
- Their use in regulatory validation processes is still exploratory. The European Medicines Agency (EMA) has begun to consider whether synthetic data can replace real data in clinical evaluation, especially in rare diseases. But the validation frameworks aren’t fully defined. Assuming a synthetic set carries the same regulatory weight as a real one is a mistake that can compromise the approval of a product or process.
What your architecture needs to integrate them well
Generating valid synthetic data isn’t hard with today’s tools. What’s complex is integrating it into a real data architecture so that it’s manageable, traceable and auditable. This is the step most organizations underestimate.
The process begins before generating anything. You need to define precisely what the synthetic data will be used for, which statistical properties must be preserved and what level of fidelity is needed for that specific use case. Without that prior definition, it’s common to produce sets that pass the technical validations but aren’t useful for the real problem.
Once generated, synthetic data requires a rigorous validation process: confirming that the relevant distributions have been preserved, verifying that it contains no information that would allow re-identifying individuals from the original set, and checking that a model trained on it performs comparably to one trained on real data. That validation isn’t a final step: it’s part of the data lifecycle.
Within the architecture, synthetic data needs a defined place: how it’s stored, how it’s versioned, how its provenance is documented and how traceability is maintained between the synthetic set and the real data that originated it. Without that governance layer, the organization ends up with data of uncertain origin and unknown validity, which is exactly the problem synthetic data should have solved.
How BertIA approaches synthetic data integration
At BertIA, we work with organizations in sectors such as the pharmaceutical industry, logistics and retail, among others, that face this problem recurrently: they have AI projects that stall because real data isn’t available in sufficient volume, because regulation blocks its use in development environments, or because sharing it with external teams creates privacy risks that aren’t acceptable.
Our way of approaching this challenge always starts from a quality diagnosis of the source data. Before designing any synthetic generation strategy, we assess whether the real problem is one of volume, access or quality, because the technical answer is different in each case.
When synthetic data is the right answer, we define the generation method, the validation criteria and the governance architecture needed for that data to be a manageable asset, not a set of files of uncertain origin.
Conclusion
Synthetic data is a tool with real usefulness in very specific situations: structural data scarcity, regulatory or privacy restrictions, and development and testing environments. Outside those situations, its application requires careful evaluation, because it doesn’t fix quality problems at the source, it doesn’t always replace the variability of real data, and its regulatory weight in formal validation processes is still limited.
What separates the organizations that extract value from this technology from those that don’t isn’t the generation tool they use: it’s having first fixed the quality of the source data, defined the use case precisely and built the governance needed for the synthetic data to be traceable and auditable.
If your organization is evaluating whether synthetic data can unblock projects that are stalled by data restrictions, talk to our team to review your specific situation.
