# Learning interactives (Google Research) Learning interactives are custom, guided STEM simulations that a teacher requests and an AI model builds on the spot. Google Research presented the experiment on 17 September 2026 (Gal Elidan and Yael Haramaty), with a tech report on arXiv (2609.20738). The model generates the whole interface instead of filling a pre-coded one; that is what generative UI means. ## Why it exists Most digital learning is still passive: you read a page or watch a video. Practice works better, and that's an old idea. [[John Dewey]] wrote in 1916 that we should "give the pupils something to do", [[Jean Piaget]] showed that learners construct knowledge, and the ICAP framework (Chi and Wylie, 2014) finds that interactive work produces deeper understanding and longer retention than listening or reading. In short, [[People learn less effectively through knowledge transfer|people learn by doing]]. Interactive simulations are expensive to build, though, so there are few of them and teachers can't adapt them to a class. One high school teacher summed it up: "it's just, you get what you get." Learning interactives build on LearnLM ([[Google]]'s education-tuned models, 2024) and Learn Your Way (a 2025 experiment that rebuilt textbooks with [[Generative AI (Gen AI)|generative AI]]). ## How it works 1. The teacher gives a topic 2. The system proposes learning objectives aligned with the curriculum; the teacher approves or edits them 3. The system generates the levels, the scaffolding and the UI 4. Self-correcting loops check the result and regenerate until every criterion passes 5. A teacher validates the final interactive before it goes into the public library The design rests on three principles: objectives approved by the teacher, inquiry-based learning (students explore and test hypotheses), and factual accuracy. Each interactive works like a game ([[Gamification]]): levels get harder step by step, and each level maps to an objective. The scaffolding guides students instead of handing them answers. It includes a priming intro that activates prior knowledge, a toolbox of formulas, several tiers of hints, feedback on why a specific answer works or fails, and a worked solution once the student has tried. The quality loops cover three areas: - **Pedagogy**: do the levels cover the objectives, and does the difficulty actually increase? - **Mechanics**: do the buttons work, and can each level be solved? - **Visuals**: are there distracting objects on screen? Some evaluators are agents. The solvability check opens Chrome and plays the simulation like a student would, and it also tries adversarial moves, such as pushing every knob to an extreme value. Google says plainly that the loops slow generation down and buy higher quality. ## What's available and what the evaluations say - A public library of 30+ English interactives for middle and high school (physics, chemistry, biology, math). All AI-generated, all reviewed by teachers. Examples: Kepler's laws, projectile motion, data visualization - Schools on Google Workspace for Education can request custom interactives through the Google for Education Pilot Program - UK STEM teachers rated 40 interactives against 10 criteria. Most came out good or excellent; physics and chemistry fit the format best - In a US study, 12 teachers each requested 3 interactives and rated them 8/10 on average. They praised the ability to differentiate, the fit with their objectives, and tiered hints that resemble the help they give students one on one ## Limitations - Small, early evaluation. There's no learning-gain data yet; Google plans field studies for that - English and STEM only - Generation is slower because of the loops - Every interactive still needs human review before publication ## My take I care less about the education angle than about the architecture underneath. It's generate-then-verify: a model produces an artifact, [[Agentic loops|agentic evaluators]] attack it against explicit criteria, the loop repeats until everything passes, and a human with domain knowledge signs off at the end ([[Human-in-the-Loop]]). That pattern works for any domain where you can write down what "good" means and test it automatically. The more of the quality bar you can check by machine, the more you can trust the generator ([[AI Verifiability]]). The teacher's quote points to a bigger shift. Software used to be built once and shipped to everyone; here it's generated per need, per class, per objective. That's the same direction as [[Malleable Software]] and [[Software for One]], applied to a classroom. Once the model can build the thing, the questions that matter are who defines the criteria and who signs off. Google gave both jobs to the teacher. I think that's the right call. ## References - Google Research blog: https://research.google/blog/the-future-of-practice-enabling-teachers-to-create-learning-interactives-with-generative-ui/ - Tech report (arXiv 2609.20738): https://arxiv.org/abs/2609.20738 - Learning interactives library: http://research.google.com/p/learning-interactives - Generative UI (Google Research): https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/ - Learn Your Way (Google Research): https://research.google/blog/learn-your-way-reimagining-textbooks-with-generative-ai/ - LearnLM: https://cloud.google.com/solutions/learnlm - ICAP framework (Chi and Wylie, 2014): https://www.tandfonline.com/doi/abs/10.1080/00461520.2014.965823 - Google for Education Pilot Program: https://edu.google.com/pilot/ ## Related - [[Human-in-the-Loop]] - [[Agentic loops]] - [[AI Verifiability]] - [[Malleable Software]] - [[Software for One]] - [[Deliberate Practice]] - [[AI Evaluation]] - [[Education (MoC)]] - [[Teaching (MoC)]]