AI tutoring’s gains depend on the test

AI tutoring helped pupils answer closely related maths questions in Google and Eedi’s expanded randomised UK trial. Their 29 October technical report covers 1,179 pupils in ten state schools. A human reviewed every AI message before delivery.[1]

After an initial mistake, adjusted success on a similar question was 56.5–56.9% in the AI groups, against 43.4% with standard hints, videos and exercises. Among 916 pupils completing both Star Maths assessments, neither AI group established an overall advantage over that comparison group.[1][3]

Star Maths casts a wider net. Renaissance describes an assessment spanning number operations, algebra, geometry and data analysis. Its examples include fractions, equations, area, volume and interpreting graphs. Schools can use it at intervals to establish starting points and review attainment across the year.[2]

The breadth of those tasks explains why a general assessment asks more than whether someone can handle the next version of a practised problem. For a school choosing how to measure progress, the scope of the test is part of the question being answered.[2]

The trial covered up to ten instructional weeks; it cannot settle longer-term learning. The report’s detailed appendices are still to come. Its immediate gains warrant attention alongside the broader result.[1]

Report an error or suggest a correction