- Synthetic data is generated by models or simulations; licensed data comes from real people and systems.
- Synthetic data is cheap and scalable but can only reflect what the generating model already knows.
- Training repeatedly on model-generated data can degrade models, a problem researchers call model collapse.
- MIT Technology Review reported in 2026 that no public tests yet compare real-world data against synthetic alternatives directly.
Where synthetic data works
- Filling rare cases in a distribution you understand well
- Privacy-preserving stand-ins for testing pipelines
- Simulations for robotics where physics is well modeled
- Augmenting labeled sets with variations
Where it falls short
- It cannot add knowledge the generator lacks
- It reproduces the generator's errors and biases
- It lacks the messiness of real work that agents must handle
- Repeated training on generated data can degrade quality (model collapse)
The evidence gap
MIT Technology Review reported that no public tests yet compare training on real-world workspace data with synthetic alternatives, so the performance benefit of buying real data is still unproven publicly (MIT Technology Review). Buyers are betting it matters, which is why they pay for it.
Questions
Will synthetic data make licensed data worthless?
Not for data that captures knowledge and behavior models do not already have. Generic text is most exposed; specialist real-world data least.