Multiple choice questions (MCQs) are widely used to assess students. Motivated by issues with accuracy and reliability that were found during university exams, we conducted a controlled user experiment with 53 participants and a commercial MCQ system that used an AI engine for grading. Each participant filled in three paper answer sheets to a prescribed pattern, one with a black pen, and the others with heavy and light pencil shading. The pattern contained 100 questions (an equal number with one, two, three, four and five correct answers). The sheets were digitized using two scanners, with each set of scans graded separately and producing a similar pattern of results. In the pen condition, the AI engine did not make any grading errors and was uncertain for 0.8% of answers (those needed to be graded by hand). However, the AI engine made grading errors for 0.25% of the heavy pencil answers and 4.9% of the light pencil answers, and was uncertain for many more answers. The results show that AI grading was only reliable when participants used a pen, which raises concerns about the guidance some organizations provide for students to use a pencil. From an explainable AI perspective, conducting rigorous user evaluations would improve transparency about AI products for end-user stakeholders, help AI developers understand the limitations of their models and identify checks and balances that should be incorporated.