AI trading assistants, based on statistical patterns in historical data, tend to systematically underestimate tail risks when confronted with structural breaks and extreme black swan events — a fundamental constraint of statistical learning methods that cannot be eliminated through engineering means.