Google Launches AI-Generated Classroom Interactive Exercise Tool
Google Research has announced a study and tool called "Learning Interactives," which uses generative AI to help teachers quickly create customized classroom web interactive simulations. Teachers simply need to specify the subject, grade, and teaching topic, and the system will generate interactive exercise web pages complete with learning objectives, leveled content, layered hints, and feedback. The project is led by Google Research scientist Gal Elidan and product manager Yael Haramaty, with a related blog post published on September 17, 2026.
The entire generation process is divided into four steps: the first step involves AI generating specific, editable learning objectives based on the teacher's proposed topic, which the teacher reviews and confirms; the second step breaks down the exercise content into a leveled structure based on the approved objectives; the third step automatically generates supporting knowledge points, formula toolboxes, layered hints, contextual feedback, and complete solutions; the fourth step introduces an iterative verification mechanism with teaching safeguards, which checks the teaching logic, interaction mechanisms, and visual presentation for reasonableness through self-correction cycles.
The verification phase also includes an agent-based automated testing method: the system automatically opens a Chrome browser instance to simulate a student's normal problem-solving path and "adversarial" actions (such as random clicking, skipping levels, or entering incorrect answers) to test whether the generated interactive exercises are truly solvable and whether they encounter errors or freeze under abnormal operations, effectively conducting an automated usability quality check on each generated content.
Throughout the process, teachers maintain key veto power: from initially proposing needs, reviewing and modifying learning objectives, to the newly generated interactive exercises needing to be manually reviewed and approved by teachers before being officially included in the public exercise library for broader use. This "human-in-the-loop" design aims to prevent generated content from containing factual errors or not aligning with teaching logic from directly reaching the classroom.
The research team invited 12 U.S. teachers to participate in practical testing, with each teacher proposing three customized interactive exercise requests based on their classroom needs. The average score for the generated results was 8 out of 10. One participating high school science teacher reported that it was previously difficult to adjust teaching simulation content for different classes, stating, "You use what you have," while this tool allowed her to generate simulation content that fits her classroom progress on demand for the first time. The team also invited STEM teachers from the UK to evaluate a batch of generated interactive exercises against 10 teaching and design standards, resulting in an overall rating of "good to excellent," with the generated simulations for physics and chemistry being deemed most aligned with teaching needs.
From an industry impact perspective, such tools directly disrupt the previously labor-intensive and costly process of designing interactive teaching content. Previously, if teachers wanted interactive simulations for specific knowledge points, they often had to rely on fixed question bank-style resources, using what was available without the ability to customize for class differences. This tool shifts the production cost of customized content from professional educational content teams to the immediate requests of individual teachers. The beneficiaries are ordinary public school teachers and students who lack the budget to purchase customized teaching software; the pressure is on educational technology suppliers that provide standardized interactive courseware and question banks, whose content differentiation advantages may be diluted.
The public exercise library currently features over 30 English interactive exercises covering physics, chemistry, biology, mathematics, computer science, and earth science, aimed at middle and high school levels. Schools using Google Workspace for Education can apply for customized interactive exercises through the "Google for Education Pilot Program," with new generated content also requiring teacher approval before entering the public library.
Source: Public Information
ABAB AI Insight
Google's layout in educational technology has a clear historical context: in May 2024, it launched the LearnLM model family for educational scenarios, followed by the rollout of the "Guided Learning" mode for students in Gemini in the first half of 2025 (competing with similar products' tutoring features), and in the second half of the same year, it added interactive image deep learning capabilities to Gemini. The "Learning Interactives" is another extension of Google's continuous investment in educational vertical scenarios around LearnLM over the past two years, with the path being to first build an education-specific base model and then gradually package the model's capabilities into specific tools that teachers and students can directly use.
Google is not mobilizing traditional funding acquisitions here, but rather prioritizing research and engineering resources: it is validating the latest "generative UI" capabilities and agentic automated testing capabilities (automatically opening a browser to simulate student operations) in educational scenarios first. The motivation behind this is to treat education as a high-value testing ground for generative UI technology—educational content has high correctness requirements and low tolerance for errors, which can compel the team to refine self-verification and safeguard mechanisms more robustly. These capabilities are likely to be reused in Gemini's generative interface capabilities in other professional scenarios (such as medical and legal consulting tools).
A comparable historical case is PhET Interactive Simulations—led by the University of Colorado Boulder, it has long provided a fixed question bank for free to teachers worldwide and has been a representative of "standardized interactive teaching content" for nearly two decades, but its content updates are slow and cannot be customized for classes. Google's approach essentially upgrades the PhET-style "pre-fabricated question bank" model to a "teacher instant customization production" model. The current industry is in the early stages of transforming educational technology from a "content distribution platform" to a "content production tool."
This change is essentially a technological replacement: previously, creating a high-quality classroom interactive simulation required collaboration among educational content teams, UI designers, and front-end engineers for weeks, which was a product of professional division of labor; now, generative AI compresses the entire production chain of "text planning—level design—code implementation—teaching verification" into a single model call plus manual review. Mechanically, this is equivalent to encoding the implicit knowledge originally borne by professional teams (how to design effective layered hints, how to make interactions both engaging and aligned with teaching objectives) into the model's generation process and safeguard rules—this replacement can only be established if Google first uses dedicated models like LearnLM to make the implicit knowledge of pedagogy explicit and structured; otherwise, a purely generic large model would struggle to ensure the teaching effectiveness of generated content.
ABAB News · Cognitive Laws
The barrier is not knowledge, but the process has not been disassembled.
Whoever encodes implicit experience can replace the expert.
Customization costs are shifting from teams to individuals.