AI glossary · Basic · Data & evaluation · Language models
Synthetic data
Synthetic data is data generated artificially by an AI model or program, rather than created or collected by people in the real world. It is mainly used to train or evaluate AI models.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Foundations & applications
- Fields
- Data & evaluation, Language models
- Added
- 2026-10-08
How it works
The most common approach is to give a strong LLM examples and instructions and have it produce large amounts of data such as question-answer pairs, conversations and worked problems. Training a small model (the Student model) on data produced by a large model (the Teacher model) in this way is also a form of distillation. The generated data is not used as is; other models or rules check its quality and filter out incorrect or duplicate items. In fields where answers can be checked automatically, such as math and coding, a common approach is to keep only data with correct answers. In image, speech and robotics, data is also generated with graphics engines or simulators.
Why it matters
High-quality text on the internet is limited, so new data sources are needed to make models larger and train them better. Having people create data and label answers takes a lot of money and time, while synthetic data can be produced much more cheaply and quickly. It can also deliberately fill in rare cases or data for specific languages or domains that are hard to obtain in reality. Using fake data with similar characteristics instead of real data containing personal information can also reduce privacy risks.
Good to know
Synthetic data can inherit the mistakes and biases of the model that produced it. There are concerns that repeatedly training only on model-generated data can reduce diversity and degrade quality, so it is often mixed with real data. That is why how well you filter matters more than how much you generate. Also, some AI services prohibit, in their terms of use, using their model outputs to train competing models, so you should also check the terms of the model that generated the data.
Example
Companies building small language models (SLMs) often have a large model write large amounts of textbook-style explanations or problems with step-by-step solutions, then filter them and use them as training data for the small model. In autonomous driving and humanoid robot development, simulators are used to create varied weather, lighting and object layouts to expand training data, together with the Sim-to-Real approach of transferring what was learned to real environments.