What This Research Area Covers
Synthetic data research studies methods for generating artificial training data — often using AI models themselves — to supplement or replace real-world data, particularly useful when real data is scarce, expensive, or privacy-sensitive.
Why It Matters
High-quality training data is a major bottleneck for improving models; synthetic data offers a way to generate additional training signal at scale, though it introduces its own risks around quality and potential feedback loops.
Current Research Directions
Techniques for ensuring synthetic data genuinely improves rather than degrades model quality, methods for generating diverse and realistic synthetic examples, and understanding risks around models training on their own or similar models' output are active areas.
Related Pages
Frequently Asked
Why use synthetic data instead of real data?
Real data can be scarce, expensive to label, or privacy-sensitive; synthetic data offers a scalable alternative, though with its own quality risks.
What's the risk of training on synthetic data?
If not carefully managed, models can degrade in quality from training on their own or similar models' output, sometimes called model collapse.
Is synthetic data used in training current frontier models?
Yes, increasingly, often as a supplement to real data rather than a full replacement.
Where can I learn more about datasets generally?
See our Datasets topic page.