Synthetic Data
Definition
Synthetic data is training or evaluation data generated by a model or an algorithm rather than collected from the real world or produced by human judgment.
Synthetic data is useful because it is cheap, fast and avoids some privacy issues. It works well for augmenting datasets, covering rare cases and bootstrapping early training.
Its main limit is that a model cannot generate knowledge it does not have. For hard, high-stakes or frontier questions, synthetic data tends to repeat the generating model's blind spots, and training on it can amplify existing errors. This is why synthetic data is usually combined with, not a replacement for, real human expert judgment on the difficult edge cases.
The practical pattern is a mix: synthetic data for volume and coverage, and human expert data for the judgments that decide whether the model is actually correct.
How Pathwize relates
Pathwize supplies the human expert judgment that synthetic data cannot reach: the hard edge cases where being correct depends on real domain knowledge.
Frequently asked
Data generated by a model or algorithm rather than collected from the real world or produced by human judgment.
Not for hard or high-stakes judgments; a model cannot generate expertise it lacks, so human expert data is still needed on the difficult cases.
For adding volume, covering rare cases and early training, usually combined with real expert data on the hard questions.
Related terms