The concept of 'Calibration' in AI safety refers to a model's ability to accurately represent its own confidence levels, including acknowledging uncertainty.