For " edge cases " or " rare scenarios " (often called " black swan " events), historical data is insufficient because the events haven ' t happened yet. " Synthetic data " is artificially generated data that maintains the statistical properties of real data but includes simulated rare events. This allows auditors to " stress test " how the AI model would react to extreme conditions, such as a massive market crash or a rare medical emergency.
Anonymized or de-identified data only reflects past occurrences. Using synthetic data is a proactive testing strategy to ensure the AI ' s safety and robustness in unpredictable environments.