Guide
Synthetic data vs licensed real data for AI training.
When data generated by algorithms works for training AI models, when real data licensed from the companies that created it is the better choice, and how the two are combined.
WhatsApp and Telegram +44 7563 115980
[email protected]
When data generated by algorithms works for training AI models, when real data licensed from the companies that created it is the better choice, and how the two are combined.
Since 2013 we have helped firms with due diligence and data services, in more than 60 countries.
The short answer
Synthetic data is artificial data produced by an algorithm to copy the patterns of real data. It is useful for filling gaps, creating rare cases and testing, but it is only as good as the real data and models behind it. Research published in Nature in 2024 found that training models on content generated by other models, without care, causes irreversible defects. Synthetic data is not automatically anonymous, carries the biases of its source and must be disclosed under the EU AI Act. Licensed real data provides the ground truth that synthetic data cannot invent, so most strong training pipelines anchor on real data and use synthetic data to extend it.
On this page
The ICO (Information Commissioner's Office, the UK privacy regulator) describes synthetic data as artificial data generated by data synthesis algorithms, which replicates the patterns and statistical properties of real data. It is generated from real data using a model trained to reproduce its characteristics and structure.
Licensed real data, by contrast, is data created by people and systems in the course of real activity, obtained under a written licence from its owner. Our AI training data service supplies data of this kind.
| Question | Synthetic data | Licensed real data |
|---|---|---|
| Where it comes from | An algorithm trained on real data, or an AI model's outputs | People and systems doing real work, licensed from the owner |
| Realism | Reproduces patterns the generator learned; can miss rare and unexpected cases | Contains the full variety of real behaviour, including the rare cases |
| Volume | Can be produced in large amounts once a generator exists | Limited to what exists, though it may keep growing |
| Privacy | Not automatically anonymous; depends on whether source data can be inferred | Must be de-identified and tested before release |
| Bias | Carries through the biases of the source data | Reflects the real population it came from, which must be checked |
| Rights | Depends on the source data and the terms of the model that generated it | Cleared and recorded in the licence |
| EU AI Act summary | Section 2.5: disclose synthetic data and the model used to generate it | Section 2.2.1: report that licensing agreements exist and the modalities covered |
| Best use | Filling gaps, rare cases, testing, balancing classes, privacy where real data cannot be shared | Grounding models in reality, evaluation, expert and human behaviour, domain knowledge |
In July 2024 researchers published a study in Nature showing that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which the tails of the original content distribution disappear. They called this model collapse and showed it can occur in large language models and other generative models.
The tails are the less common cases: unusual phrasings, rare events, minority views and edge conditions. They are often exactly what makes a model reliable in real use. When each generation of a model trains on the output of the last, those cases thin out and then vanish.
The authors concluded that data collected about genuine human interactions with systems will be increasingly valuable as more model-generated content appears online. That is the strongest argument for keeping licensed real data at the core of any training set. Our guide to AI data licensing deals covers how that value is reflected in the market.
We supply licensed company datasets, rights cleared and de-identified, to use on their own or to ground and test synthetic data.
Speak to an expertAI training dataSynthetic data is often presented as a privacy solution. The ICO's guidance is more careful.
Where synthetic data is generated from personal data, the same tests apply as for any de-identified release. Our guide to anonymisation vs pseudonymisation sets them out.
Synthetic data does not escape questions about where it came from. Two points matter for providers of general-purpose AI models in the EU.
Licensed real data is reported more simply. For data licensed from rights holders, the template asks only whether such agreements exist and which types of content they cover. See our EU AI Act training data summary guide.
The choice is rarely one or the other. The strongest approach uses real data as the anchor and synthetic data to extend it.
Synthetic data tends to work best where the real data already exists but is scarce in one area, and the generator can be checked against real examples. It works least well where the task depends on judgement, expertise or behaviour that the generator never saw, such as how specialists handle unusual cases. In those areas only real records carry the information.
Data owners whose records could anchor a model can read our guide on how to sell data to AI companies.
Artificial data generated by an algorithm to replicate the patterns and statistical properties of real data. It can be partially synthetic, replacing some fields, or fully synthetic, replacing all of them.
Not on its own. Synthetic data is useful for filling gaps, rare cases and testing, but it can only reproduce what its generator learned. Licensed real data provides the ground truth and is essential for evaluation.
Not automatically. The ICO says it depends on whether the source personal information can be inferred from it, and some generation methods are vulnerable to attacks that recover source data or reveal whether a person's record was used.
A loss of quality, identified in a 2024 Nature study, that occurs when models are trained indiscriminately on content generated by other models. The less common cases in the original data disappear, and the defects are irreversible.
Yes. The training data summary template asks providers of general-purpose AI models whether they created AI-generated data for training and which models generated it.
It can reduce how much real data is needed for some tasks, but it cannot replace the real cases it was never shown. Models trained without fresh real data risk losing accuracy on rare and unexpected situations.
No. It inherits questions about the data used to train the generator, and the terms of the generating model may restrict how outputs are used. Its origin should be recorded like any other source.
Compare models trained with and without it on held-back real data the generator never saw, check that rare cases survive, test for bias, and test privacy where the source contained personal information.
To fill specific gaps in real data, create rare or sensitive scenarios, balance under-represented cases and test systems. Use it alongside licensed real data, generated from cleared sources, and always evaluate the model on real data it has not seen.
Each point of law and each reported figure on this page was checked against these sources on 27 October 2026. This page is general information, not legal advice.
Tell us what your model needs to learn and the data you have today.