Noninvasive blood glucose measurements are characterized in human subjects. A series of first overtone transmission spectra are collected across the tongues of five human subjects with type 1 diabetes. The noninvasive human spectra are collected by an experimental protocol that is designed to minimize chance correlations with blood glucose levels. In one treatment of the data, every fifth sample is used as a blind prediction point to validate model performance. In another rearrangement of the data, the spectra collected over the first 29 days are used to build calibration models that are then used to predict in vivo glycemia from spectra collected over the next 10 days. Of the five data sets (one for each subject), one demonstrates a complete inability to predict blood glucose levels and is deemed void of glucose-specific information. Glucose-specific information is evident in the remaining four data sets, albeit to varying degrees. For all data sets, the ability to measure glucose from spectra collected noninvasively from human subjects depends on spectral quality and reproducibility of the tongue-to-spectrometer interface. The standard error of prediction is 3.4 mM for the best calibration model. The significance of this magnitude of prediction error is discussed relative to the situations where: (1) the model is completely void of glucose-specific information and (2) glucose predictions are limited by spectral signal-to-noise and sample thickness. Overall, glucose-specific information is available from noninvasive first-overtone spectra collected across human tongues. Significant improvements are necessary, however, before clinically useful measurements are possible.