synthetic data

in English: synthetic data

basicspreading

Artificially generated data that mimics the characteristics of real data but does not originate from real people or events.

In plain terms

Synthetic data is like a flight simulator for pilots: it is not real flight, but it is realistic enough to learn from. Here, the machine does not work with real data from real people, but with invented yet realistic data: records of non-existent patients, bank card transactions that never happened, and staged traffic situations.

There are three main reasons for its use. The first is data protection: real patient data or bank transactions cannot simply be handed over to a development team. The second is scarcity: there is little real footage of rare diseases or rare accident types, but any amount can be artificially created. The third is cost: collecting and labeling real data is expensive and slow.

It differs from real data in that it is generated rather than observed. Therefore, it is only as good as the underlying rule or model's understanding of reality. What the creator does not know about the world will also be missing from the synthetic data.

How it works

First, a starting point is needed: a real dataset whose patterns can be learned, or a rule system, simulation, describing how the world behaves.
If starting from real data, a generative model learns its characteristics: what values occur, how often, and what correlates with what (for example, age with blood pressure). In simulation, this role is filled by built-in rules.
The model or simulation then generates new records, texts, or images that never existed before. Collectively, these resemble the original, but none is a copy of a real case.
The output is checked: is it realistic enough, can it be used for the same purpose, and can real individuals not be traced back from it?
The prepared data is used for training, testing, or sharing. It is often mixed with real data so the system does not detach from reality.

An everyday example

Developers of a banking application do not test the new feature on your transaction history but on thousands of fictional transactions from fictional customers. Similarly, the software for autonomous vehicles practices countless dangerous scenarios in simulation that would be life-threatening to arrange on real roads.

Who is affected

developerWithout real customer data, it can test systems, train models or agents, and fill in rare cases. → Code

patient, doctorHealthcare research can progress without real patient records leaving the hospital. → Health

bank customer, investorFraud filters and credit scoring systems are increasingly trained and tested on artificial transactions as well. → Money

lawyer, citizenSynthetic data can reduce data protection risks. However, if it is generated from real personal data, the generation process itself falls under the scope of the GDPR. The result only falls outside the GDPR if it is proven that real individuals cannot be inferred from it. → Law

A common misconception

MythSynthetic data is fake, so it is worthless, or conversely: since it is made up, it is automatically anonymous and secure.

In realityNeither is true: well-prepared, it is useful teaching and testing material; poorly prepared, it is misleading. If it follows the original too closely, data of real individuals may leak from it.

Limits and risks

It inherits errors and biases from the source data and can even amplify them.
It can reproduce truly unexpected, never-before-seen situations only to a limited extent, because it is built on what its creator or the source data already knew.
If an AI model is trained predominantly on data generated by other models, its quality may degrade from generation to generation. This is called model collapse.
The fact that data is synthetic does not prove it is anonymous; this must be examined and verified separately.

Similar, but not the same

anonymized data – Anonymized data is real data from real people from which identifiers have been removed. Synthetic data, by contrast, is newly generated.
deepfake – A deepfake is typically a fake image, audio, or video of a real person, capable of deception. Synthetic data, however, is mostly created to train and test machines, not to deceive people.
data augmentation – Data augmentation slightly modifies existing real data (for example, rotating an image). Synthetic data creates entirely new examples.

Related concepts

Broader

training data, adat, machine learning

Narrower

synthetic training data, simulation data, szintetikus betegadat

On the edge

generative AI, data augmentation, anonymization, model collapse, digital twin

A brief history

The concept of synthetic data in the modern sense was proposed by statistician Donald Rubin in 1993, to allow census data to be published without revealing respondents' identities. Generative models, especially GANs introduced in 2014, and then large language models, made it possible to mass-produce lifelike images and texts this way.

Where things stand

As of late 2026, synthetic data is an established tool for training AI models and for software testing, and it is also spreading in healthcare, finance and the automotive industry. Unified standards for verifying quality and anonymity are still taking shape.

Status: 2026 November

Also known as

szintetikus adat · szintetikus adatok · synthetic data · artificially generated data · synthetic data · synthetic dataset · artificial data

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.