Google tests AI-made classroom simulations. Learning gains remain unproven

A simulation built around a lesson

Google Research described a system on September 17 that lets teachers shape interactive learning activities with AI. Its sample library offers more than 30 English-language STEM activities for middle and high school, reviewed by teachers.

The approach uses generative UI: AI produces an interface students can manipulate, rather than only a written answer. Teachers approve learning objectives, then the system builds guided tasks with feedback, hints and worked solutions. Automated checks try to catch problems with both the interface and the task.

The attraction is easy to see. A teacher could want practice fitted to a particular lesson, not a general explanation of the topic. The harder question is whether the generated activity teaches the right idea.

A polished example has a production history

The technical report gives a useful view behind the demonstrations. Starting from validated simulation ideas and level goals, 3.5% of generated simulations passed all automated critiques on the first attempt. After ten improvement rounds, the figure was 69.3%. These are generation-check results, not student scores.

In a separate usability study, 12 US STEM teachers supplied 36 requests. Researchers generated five candidates per request and had internal raters select the best. Three requests produced nothing of sufficient quality to show.

Those choices matter when interpreting positive teacher feedback. The assessment concerns selected, refined examples. It does not tell a school how often a teacher working alone would obtain an equally useful activity, or how much correction time the process would take.

Working controls are only the first test

Our reading is that three different checks are needed. Does the activity work? Does it represent the subject accurately? And does using it help a pupil understand something they could not explain before?

A simulation can pass the first check while leaving the other two unresolved. Attractive motion is not evidence of understanding. Nor is a teacher finding an interface easy to use the same as a student retaining the lesson.

The report identifies classroom learning effectiveness as work still to do and describes a planned pilot. The sensible next evidence would compare learning outcomes and account for teacher preparation time, rather than relying only on approval of the interface.

How access works for now

Google directs interested schools to its Education Pilot Program. That page says a Workspace for Education administrator must register interest; an invitation follows when a suitable pilot is available. Registration is not immediate access to every experiment.

The linked activity library redirected our unauthenticated access check to Google sign-in. We have not independently tested its exercises. For now, this is promising work on making practice materials, with the educational benefit still to demonstrate.