Synthetic Data

Synthetic data is training or evaluation data generated by a model rather than collected from the real world. It's used to cover cases real data is missing, scarce, expensive, or too sensitive to use — and increasingly to train models when high-quality human data runs short.

Also known as: synthetic training data, generated data

Jun 27, 2026 · Chain of Thought

Real data is the bottleneck in most AI work: it’s expensive to label, scarce for rare cases, and often can’t be used at all for privacy reasons. Synthetic data sidesteps that by generating examples — a stronger model writing training pairs, simulated edge cases, augmented variations — to fill the gaps. DeepSeek’s use of reinforcement learning to auto-generate step-by-step reasoning data is a high-profile example of synthetic data shaping a frontier model.

The leverage is real, and so is the risk. Synthetic data inherits the biases and blind spots of the model that made it, and training a model heavily on another model’s output can amplify errors or narrow diversity. The practical stance builders take: use it to augment and cover gaps, validate it against real data and good evaluation, and don’t mistake volume for quality.

Go deeper

Wang et al.: Self-Instruct paper

Follow the instruction, input and output generation and filtering pipeline, then inspect the expert-written tasks used for evaluation. Keep the instruction-tuning setting in view when adapting the method.

Shumailov et al.: The Curse of Recursion paper

Inspect the recursive-training setup and distribution-tail behavior in the experiments. Use the stated conditions when assessing a synthetic-data pipeline.

Synthetic Data

Hear it from the guest

Go deeper

From the conversation

Synthetic Data

Hear it from the guest

Go deeper

From the conversation

Related terms