New Audit Tool Checks If AI Tutors Teach or Give Answers

AI programming tutors are popular for people learning to code. But many simply provide model answers, which feels helpful while secretly robbing you of practice and understanding. A new research paper introduces ExplainRoute, a pre-deployment audit framework—essentially a safety checklist that inspects a tutor before it is allowed into a classroom. It checks whether the tutor is a good coach or just an answer machine.

The framework looks at a line of code and a learner's own explanation of it. Then it decides between two response styles: a Feynman-style prompt, asking the learner to explain the code as if teaching a child, or a Socratic scaffold, which guides with thoughtful questions. Crucially, ExplainRoute writes down everything it does—including the selected strategy and possible leaks of the answer—in a machine-readable contract. It was tested on 1,770 pairs of real code and learner explanations, with a held-out portion to prevent the system from simply memorizing examples.

The results are striking. Complex adaptive routing—choosing a different prompt strategy each time—was not better than using one fixed, proven strategy. An independent evaluator gave adaptive responses 4.52 out of 5, but simple Socratic scaffolding scored 4.66. The adaptive system only matched the rule-based reference 60.5% of the time. In other words, all the extra intelligence did not add teaching value. The same findings held in a blinded rubric review on a smaller sample.

For students, this means a less glamorous, question-based tutor may be more honest and effective than one that sounds "smart." For developers and schools, it highlights the need for independent audits before educational AI is deployed. The researchers' contribution is not proof that students learn more—it is a reliable way to verify that AI tutors respect the boundary between supporting and spoiling a learner. The bottom line: everyday learners deserve AI that makes them think, not just AI that produces answers.