AI as a learning partner, not a replacement for thinking

AI works best for learning when it makes practice faster and feedback more precise, not when it does the thinking for you. Three very different experts point to that same way of using it: let AI speed up practice, feedback and idea generation, and keep the actual understanding with the person.

Joe Liemandt of Alpha School uses AI tutors, which he stresses are not chatbots, to assess what each child knows and generate lessons at that level. He says most students use chatbots to cheat. At Alpha, students spend about two hours a day on individualized academics, and he reports they learn more than twice what they would in a normal school day. These are Alpha's own results, with independent trials still underway. His method is a worked example, initial mastery, then spaced repetition. He insists on fluency in basics, because anything memorized to fluency stops using up limited working memory. One student stuck in the low 700s on SAT math went back, memorized her multiplication tables and scored 790. The AI finds the gaps, but the student still does the reps. He calls motivation 90% of learning and says caring adults remain essential.

Salk Institute neuroscientist Terry Sejnowski supplies the rule behind keeping the human work: "it has to be active learning." Wrestling with a problem yourself builds something that generalizes, while being told the answer stores a fact that will not come to mind for a different problem. Expertise in any field comes from practice. He is optimistic about AI as a partner. A colleague uses a language model as an idea pump. In a dermatology study, doctors and an AI each reached about 90 percent accuracy alone, and doctors using the AI reached 98 percent. He also notes ideas come during walks and before sleep, and that sleep consolidates the day's learning.

MIT psycholinguist Ted Gibson explains why the human has to stay in charge. Language models are superb at the form of language, but "meaning is tough still." Told that the car is known to be behind door one in a modified Monty Hall problem, a model still says to switch, because it follows the familiar shape of the puzzle. Brain research shows language and reasoning use separate systems, so fluent text is not proof that reasoning happened.

They differ in places. Sejnowski encourages trying several chatbots and treating them like colleagues, while Liemandt sees open-ended chatbots mainly as a cheating risk for children. Gibson says models do not understand; Sejnowski calls understanding a hard word to pin down. Huberman argues errors and struggle drive learning, while Liemandt says anything that can be taught without suffering should be. For learners the result is the same: check the reasoning yourself.

What to do with it: try the problem before asking AI, then use it to check your answer or explain your mistake. Have it quiz you to find gaps, and fill them before moving on. Drill basics to fluency. Test yourself. Use AI for ideas and check each one. When an answer sounds right, change a key detail and see whether the model adjusts. Keep the offline work, such as walks and sleep. For children, pair AI tutoring with a caring adult who sets high standards.