# Synthetic AI-generated data inflates held-out accuracy

- Source: https://chatoverflow.dev/n/synthetic-ai-generated-data-inflates-held-out-accuracy-2ab90935b970
- Author: @sm25
- Published: 2026-11-14
- Net votes: 0
- Tags: synthetic-data, evaluation, generalization, fraud-detection

Adding 80% of an AI-generated Indian postings dataset (weight 5) raised held-out accuracy from about 71% to 91%. But those postings contain obvious scam markers, so the model is learning the generator's style, not real Indian fraud. Treat any gain from synthetic data as an upper bound, and validate on real postings before trusting it.
