Possibilities, limits and educator decisions
AI can help educators draft materials, explore explanations and prepare questions. Its value depends on the learning goal, the quality of the output and the decisions people make about using it.
Historical development
The field developed through specialist tutoring, automated assessment and conversational tools. These dated milestones show changes in research focus; they do not mark universal adoption in schools.
Good examined intelligent computer-assisted instruction, including models of the learner, teacher and natural environment. Diagnosing misconceptions and providing meaningful science-learning environments were central concerns.[1]
A later Scopus review found relatively few publications during this period. Its search begins in 2002, which is a database coverage boundary, not the origin of AI in science education.[2]
Liu and colleagues validated machine scoring for eight inquiry assessment items requiring students to explain phenomena using evidence. The study examined agreement with human scoring for these items, establishing a specific assessment application.[4]
Zhai and colleagues reviewed 49 studies of machine learning in science assessment. Much of the work concerned scoring accuracy and validity; fewer studies investigated how these systems changed classroom teaching and learning.[5]
Inquirybot used planned dialogue to guide an investigation of sound transmission. In the same year, Cooper explored ChatGPT for science explanations and teaching materials. A scripted inquiry chatbot and a general-purpose generative model offer different kinds of support.[6][7]
Jia, Sun and Looi synthesised 76 studies of primary and secondary science education, identifying applications including educational robots, tutoring and data-driven assessment. Their review was published online on 6 October 2023 and appeared in the 2024 journal volume.[3]
The review reports more publications after 2015. The 2023 bar covers only January–May, so it is not directly comparable with a complete year.
The same review shows an uneven distribution of publications across countries. This describes indexed research output, rather than how widely teachers use AI.
Original graphs reproduced from Akhmadieva et al. (2023), p. 5.[2] © The authors, CC BY 4.0. Extracted from the PDF without altering the graphs. Select a graph to view it at full size.
Before you use AI-generated science content
Keep checking part of the teaching work. This guide offers reminders; it does not certify content.
Six principles for educators
A short set of questions to ask before you place generated material in a science learning activity.
What should learners understand or do, and how will they show it? Choose an AI role only after that is clear.
Technological choices make sense alongside the science content and teaching method. The TPACK account treats these forms of teacher knowledge as connected, so begin with the learning goal and select a tool only if it helps students reach it.[8] For example, a generated question is useful when it asks students to explain evidence relevant to the lesson objective.
Plan with an objective →The educator selects the materials, guides instruction and makes assessment decisions. AI output is a suggestion to review.
Large language models can assist with preparation and feedback, but their responses can be unreliable and shaped by bias. Educators therefore decide which output is suitable, explain its role to learners and retain responsibility for consequential judgments.[9]
Browse teacher resources →Confirm important claims, equations, assumptions, units and references against course-approved material.
Fluent writing can hide mistakes. Published chemistry examples document incorrect calculations and fabricated references in ChatGPT output.[11] Rework an answer yourself and compare its claims with dependable disciplinary sources before it reaches students.
Use the quick check →Make room for a learner's own prediction, evidence, explanation and independent attempt.
Use AI dialogue to ask learners to examine and challenge an explanation, not simply receive one. A chemistry teaching study explored ChatGPT-supported critique and students' reported confidence; it does not establish that AI alone improves critical thinking.[10] Ask students to state an initial view, test it against evidence and explain any revision.
See the examples →Tell learners what help is permitted and how you will assess the scientific reasoning the activity is meant to reveal.
If a task assesses a student's explanation from evidence, specify whether AI may help with wording, planning or neither. Automated scoring research addresses selected response types and does not remove the need to interpret a student's reasoning in context.[4][5] Ask for a record of permitted assistance when relevant.
Plan assessment expectations →Do not make a paid tool, personal device or account a hidden requirement. Offer a meaningful equivalent route.
Opportunities from language models depend on access, digital skills and how systems serve different learners. The education literature raises fairness and inclusion as design concerns.[9] Provide a comparable non-AI route and check that both routes let students meet the same objective.
Plan an alternative →References
The numbered links in the text lead to the original papers and reviews below.
See the ideas in action