A Quiet Revolution in Physics Learning: Students Are Moving from Google to Generative AI
Students across the world appear to be changing the way they search for answers to physics questions, with generative artificial intelligence increasingly replacing the traditional journey through search results, educational websites and online encyclopedias. A new international study led by researchers at Bar-Ilan University in Israel has found a sustained decline in Google searches for physics concepts between 2022 and 2025, a period that coincided with the rapid public adoption of tools such as ChatGPT. The researchers emphasize that the trend does not demonstrate that students are learning less. Instead, it suggests that information-seeking itself is being reorganized: rather than entering a question into a search engine and comparing several sources, many learners may now be asking an AI system for a direct, conversational explanation.
The study, led by Dr. Yossi Ben-Zion and Omer Michaeli of Bar-Ilan University’s Department of Physics in collaboration with Professor Noah Finkelstein of the University of Colorado Boulder, examined patterns in more than 20 countries. Their findings, published in Physical Review Physics Education Research, offer one of the first large-scale attempts to measure how generative AI is reshaping the public search behavior associated with science learning. The analysis focused on physics because the subject contains a wide range of concepts that students commonly investigate independently, from force and energy to relativity and quantum mechanics. By tracking changes in online activity over time, the researchers were able to compare information-seeking behavior before and after generative AI became widely available.
The team used Google Trends, a platform that reports the relative popularity of search terms across time and geographic regions. Unlike raw search counts, Google Trends data are normalized, meaning that activity is presented in proportion to all searches conducted in a given location and period. This allows researchers to compare patterns across countries with very different populations and internet usage levels, although it does not reveal exactly how many people searched for a term or whether every search was made by a student. Across the countries included in the study, searches for numerous physics concepts declined after the emergence of widely accessible generative AI tools. To test whether the change was specific to Google, the researchers also examined Wikipedia page views for physics articles in seven languages and identified similar downward patterns.
The most surprising geographic divide appeared between English-speaking and non-English-speaking countries. In the United States, the United Kingdom and Australia, Google searches for many physics topics remained comparatively stable. In numerous countries where English is not the primary language, however, search activity fell substantially. The researchers argue that this difference may reflect more than variation in technology adoption. It may reveal a major educational advantage of multilingual AI systems: the ability to explain difficult scientific ideas in a learner’s native language. For students who have traditionally depended on English-language search results, textbooks or lecture notes, conversational AI may reduce the effort required to translate both vocabulary and meaning.
That burden, sometimes described as a “language tax,” can affect every stage of scientific learning. A student may understand the physical idea being discussed but still need to translate a question, identify the correct technical term in English, evaluate unfamiliar sources and then translate the explanation back into a preferred language. Generative AI can compress those steps into a single interaction, allowing a learner to ask a question using everyday language and request an explanation at a particular level of difficulty. The researchers suggest that this capability could make AI an educational equalizer, giving more students access to explanations that were previously easier to obtain for people fluent in English or familiar with the structure of online academic resources.
The transition from search engines to AI was not uniform across all areas of physics. Search activity declined most sharply for topics that can be explained mainly through text, definitions and logical reasoning. Concepts that depend heavily on diagrams, graphs, spatial relationships or other visual representations showed smaller decreases. This distinction provides a window into the current technical profile of large language models. Such systems are highly effective at producing fluent verbal explanations, tracing chains of reasoning and adapting their language to a user’s apparent level. They are often less reliable when a learner must interpret a complex graph, visualize a three-dimensional system, follow a vector diagram or connect several visual elements at once.
This pattern does not mean that AI cannot generate images or discuss visual material. Rather, it points to a continuing difference between language-based explanation and genuine visual understanding. Physics frequently requires students to move between equations, words, diagrams and physical intuition. A verbal description of an electric field may be helpful, but it does not necessarily replace the ability to read field lines or reason from a spatial model. Similarly, an explanation of a motion graph may fail to develop the skill of extracting acceleration, turning points or changes in slope directly from the figure. The findings therefore suggest that the subjects most easily transferred to conversational AI may be those for which language can carry much of the conceptual load.
The researchers caution that their study measures information-seeking behavior rather than learning outcomes. A reduction in Google searches could mean that students are receiving useful answers from AI, but it could also mean that they are searching less because they have stopped investigating a topic, are relying on unreliable explanations or are asking questions through platforms not captured by the analysis. Google Trends and Wikipedia page views provide indirect indicators, not a direct record of comprehension, persistence or academic achievement. AI-generated explanations can also contain factual errors, omit important assumptions or produce confident but invalid reasoning, especially in areas involving mathematical derivations and multi-step problem solving. The shift in behavior therefore raises questions for educators about how students should verify AI responses and demonstrate understanding.
For science education, the implications may extend well beyond physics. If learners increasingly bypass conventional search engines, educational institutions may need to reconsider how digital resources are designed and how research skills are taught. Search literacy traditionally involves selecting keywords, comparing sources, recognizing authority and following references. Conversational AI changes that process by presenting a synthesized response before the learner has encountered the underlying sources. This can make information easier to access, but it may also hide disagreement, uncertainty and the path by which an answer was constructed. Teachers may increasingly need to assess not only whether students reach a correct result, but whether they can interrogate an AI explanation, identify its assumptions, check its mathematics and connect its claims to evidence.
The study ultimately portrays the rise of generative AI not as a simple replacement of Google, but as a transformation in the architecture of learning. Students appear to be moving from a model based on navigating information landscapes to one based on interacting with an artificial conversational tutor. That transition may broaden access to science, particularly for learners working outside English, while exposing weaknesses in areas that depend on visual reasoning and independent source evaluation. As AI systems become more capable and more deeply integrated into education, the central challenge will be to preserve the curiosity, skepticism and conceptual discipline that physics demands. The next generation of learners may ask fewer questions of search engines, but the quality of the questions they ask—and their ability to judge the answers—could become more important than ever.
Subject of Research: Global changes in physics-related information-seeking behavior associated with the adoption of generative artificial intelligence.
Article Title: From search to GenAI queries: Global trends in physics information-seeking across topics and regions
News Publication Date: 17-Aug-2026
Web References: Physical Review Physics Education Research article; DOI link
References: Ben-Zion, Yossi; Michaeli, Omer; and Finkelstein, Noah. “From search to GenAI queries: Global trends in physics information-seeking across topics and regions.” Physical Review Physics Education Research. DOI: 10.1103/g1vh-fl9s.
Keywords
Generative AI, ChatGPT, physics education, Google Trends, Wikipedia, information-seeking behavior, multilingual education, science learning, artificial intelligence, educational technology
