Studies find AI tutoring can help students, but works best with humans

In short: Despite some criticism, research so far generally shows AI tutoring can improve learning a little, especially when it helps human tutors.

What's going on

A New York Times opinion piece argues that AI tutoring is “seductive but ultimately ineffective.” But the evidence from recent studies does not fully match that framing. Across multiple reviews and real-world trials, results tend to show small to moderate improvements in learning, not zero impact.

One of the clearest examples comes from Stanford researchers who tested an AI helper called Tutor CoPilot during live online math tutoring. In a randomized controlled trial (a study where people are split into groups by chance to compare results), students were 4 percentage points more likely to pass session assessments when their tutor used the AI assistant. The biggest improvement showed up for less-experienced tutors, whose students’ math proficiency rose by about 9 percentage points, close to the results of top-rated tutors.

Other research looks at “intelligent tutoring systems,” which are learning programs that adapt as a student works (like a workbook that changes the next question based on your last answer). A review of these systems in K-12 schools found overall positive effects, though not dramatic ones. Some gains appear after sustained use, including one algebra program where year-two students outperformed a comparison group.

There is also early evidence that AI tutoring may help more in places with fewer resources. A small after-school pilot in Nigeria reported fast gains in English skills over six weeks.

What to watch

The main question now is not whether AI tutoring ever helps, but when it helps and what risks come with it. Researchers are asking for longer studies, plus clearer rules for accuracy, student privacy, and how to keep humans in the loop so AI acts more like a co-pilot than a substitute.

Source: NYTimes