Intelligent Tutoring

The dream of a machine tutor that adjusts perfectly to every learner has been around for longer than the smartphone.

For decades, research groups have been building systems that try to do what a good human tutor does. Watch the learner, figure out where they are stuck, offer the right hint at the right moment, escalate or back off based on what happens next. The early systems were promising. The later systems were more sophisticated. The newest ones, powered by modern machine learning, can do things the earlier ones could not even attempt. Every generation of these tools has been described, in its own moment, as close to a breakthrough. Every generation has produced modest real-world gains that never quite matched the promotional material. The gap between what these systems can do in a lab and what they do in a real classroom, full of real distractions and real teenagers, has stayed stubbornly wide.

The reason is not that the technology is bad. The reason is that tutoring is not really about explaining things. A good human tutor reads the student's mood, notices when confusion is disguised as boredom, knows when to push and when to wait, recalls the cousin's wedding last weekend and how it might be affecting this week, and makes the student feel that a particular adult is on their side. A system that matches the pedagogical moves of a tutor but not the human surround is doing half the job, and half the job is sometimes worse than none, because it trains the student to perform rather than to think. The latest AI tutors will be more impressive than anything we have seen. They will still be trying to do only some of what happens between two people in a quiet room. What is the one thing your best teacher did that no piece of software could replicate?

Last week we said an AI-written quiz is a draft, not a measurement. This week we look at the long-running dream of a machine tutor and at why it keeps almost arriving.