Training
Synthetic data
Artificially generated data — from a model or simulation — used to augment or replace human-collected data when training and evaluating AI models.
Synthetic data is artificially generated data that mimics the patterns of real data rather than arising from real-world events. It is produced by algorithms, simulations, or AI models themselves, and is used to augment or fully replace human-collected datasets when training and evaluating models.
In current practice, a strong “teacher” model generates prompts, answers, or labels on which another model is trained — an approach closely related to knowledge distillation. This yields data for fine-tuning, chain-of-thought reasoning corpora, and preference pairs for RLHF. Simulations and procedural generation further fill in rare or privacy-sensitive scenarios, often with more accurate labels than manual annotation allows.
The topic is especially active in 2025–2026 as the supply of high-quality human web text runs short (the “data wall”). Research warns of model collapse when a model is trained mostly on its own outputs, so practitioners stress factuality, fidelity, and unbiasedness, and mix in real data to keep models anchored to reality.