Researchers from the University of Colorado tested the trustworthiness of several AI models with games of Sudoku. It didn't go so well.