AI systems have only limited ability to understand and reflect on their own functioning and performance limitations, presenting practitioners with challenges in assessing system reliability and underscoring the necessity of human oversight.