Skip to main content
Back to timeline
Google ResearchSource publication:

Teachers stop settling for off-the-shelf simulations: generative UI builds guided interactives per curriculum, 12 teachers rate 8/10

Synopsis

Google Research describes a research experiment in which teachers use generative UI, where a model builds the interface on the fly instead of it being coded in advance, to create guided interactive simulations for any STEM topic. Such practice activities were previously expensive to make and limited in number, leaving teachers with off-the-shelf simulations that could not be differentiated. Generation runs through self-correcting loops driven by pedagogical guardrails until objectives, solvability and visual criteria are met; 40 interactives were rated good or excellent overall by UK STEM teachers, and 12 US teachers gave custom interactives an average of 8 out of 10.

AI-generated editorial illustration: The future of practice: Enabling teachers to create learning interactives with generative UI

Interpretation

Teachers can now generate custom interactive learning simulations for any STEM topic, matched to their own curriculum goals and grade level, instead of choosing among a limited set of ready-made resources. Interactive, multimodal forms of practice were previously expensive to create, limited in number and often demanded more effort from the teacher, so they remained largely out of reach. In an initial study with 12 US teachers, each requested three custom interactives for their own classroom needs, and teachers rated the quality of the interactives 8 out of 10 on average.

The teacher supplies the topic, the model produces modifiable learning objectives that the teacher must approve, and each complex topic is broken into levels of progressively harder challenges. Generative UI is usually aimed at quick interactions, whereas this work targets deeper educational journeys and therefore needs pedagogical constraints. Each level carries an introduction, a toolbox of formulas and theories, multiple hint levels, feedback on a specific response and worked solutions, all tested and iterated with teachers and students.

Generated results pass through self-correcting loops that check pedagogy, mechanics and visual aspects, and regenerate until every criterion is met. The iterative process increases the time needed to produce the final interactive, in exchange for closer adherence to the quality criteria. Some automatic evaluations inside the loops are agentic: a solvability check opens a Chrome instance, interacts with the simulation as a user would, and tries adversarial actions such as pushing knobs to extreme values.

Perspective

What is released is a sample library of over 30 English learning interactives for middle and high school STEM, spanning physics, chemistry, biology, mathematics, computer science and earth science, all generated by AI and reviewed and approved by teachers. The intended readers are classroom teachers and instructional designers; schools can sign up through the Google for Education Pilot Program, teachers request simulations for any custom STEM concept, and generated interactives go back to the requester for review and can enter the public library only after teacher validation and approval.

Teacher ratings are good or excellent overall, with physics and chemistry most amenable to simulation creation; the 12 US teachers' feedback is an initial study with an average rating of 8 out of 10. Learning gains and student engagement remain to be evaluated through UX research and field studies, and the pilot is described as an early step toward public use.