[Paper Review] AI in Education Is Still Text-Bound: The Multimodal Gap in 4,175 Studies

A Topic-Modeling Analysis of Text-Centric Bias and Multimodal Possibilities in Generative AI Education Research

Paper Title: Beyond Text-to-Text: An Overview of Multimodal and Generative Artificial Intelligence for Education Using Topic Modeling
Authors: Ville Heilala, Roberto Araya, and Raija Hämäläinen
Conference: The 40th ACM/SIGAPP Symposium on Applied Computing, SAC ’25
DOI: 10.1145/3672608.3707764
Publication Year: 2025

The following is the complete English version of the article.

How Multimodal Is Generative AI Research in Education?

Educational research on generative AI is growing rapidly. Students use conversational AI tools such as ChatGPT to write texts, check problem-solving procedures, and generate programming code. Teachers use these tools to create instructional materials, design assessment items, and provide personalized feedback.

However, generative AI is not synonymous with large language models. It can generate or transform many different forms of information, including text, images, speech, music, video, data, tables, and graphs. Text-to-speech, which converts written text into spoken language; speech-to-text, which turns speech into written text; text-to-image, which creates images from written prompts; and text-to-video, which produces video from text, can all be used in education.

The researchers posed a straightforward question:

How evenly does educational research on generative AI address these different modalities?

The study does not test the effectiveness of a particular AI tool. Instead, it maps the broader research landscape to identify which technologies and educational themes occupy the center of current research and which remain at the margins.

Moving Beyond the Product Name “ChatGPT”

One important feature of the study is that the researchers did not use the names of specific services, such as ChatGPT, Gemini, or Claude, in their literature search.

They determined that the market influence of specific companies and products could disproportionately shape the search results. The search query therefore focused on technological concepts and transformations between modalities, including “generative AI,” “large language model,” “text-to-text,” “text-to-video,” and “text-to-audio.”

The researchers also included combinations of different information formats, such as text, video, audio, voice, sound, music, data, images, graphs, tables, emotions, language, and gestures. Their aim was not simply to identify “educational studies using ChatGPT,” but to examine how AI technologies that generate and transform different types of information are being addressed in education.

This search strategy reduces the risk of interpreting generative AI education research solely through the lens of one conversational AI service. At the same time, it reveals which technologies and interaction formats actually dominate the body of research described as generative AI in education.

Figure 1. Research Procedure Using Topic Modeling

Building a Dataset of 4,175 Publications from an Initial 4,475

The researchers collected publications from Dimensions.ai in August 2024 and conducted supplementary searches using Scopus and Web of Science. The initial search produced 4,475 records. After excluding publications without a DOI or abstract, 4,175 publications were retained for the final analysis.

The dataset included journal articles, conference papers, and book chapters. Approximately 95 percent of the publications had been published since 2014. However, the earliest relevant study dated back to 1983 and examined a text-to-speech reading aid for visually impaired students.

This finding shows that multimodal educational technologies did not suddenly emerge with the recent generative AI boom. Technologies such as text-to-speech and speech recognition have long been studied in special education and assistive technology.

Recent generative AI can therefore be understood not as an entirely new development, but as an expansion of earlier modality-conversion technologies through their integration with large-scale models and advanced generative capabilities.

Topics Generated by the Data Rather Than Predetermined by Researchers

The researchers analyzed the abstracts of 4,175 publications using BERTopic. Rather than simply counting the frequency of individual words, BERTopic transforms sentences into numerical representations that preserve their meaning and context, and then clusters documents with similar semantic characteristics.

First, the researchers used the all-mpnet-base-v2 sentence-transformer model to represent each abstract as a 768-dimensional vector. They then applied UMAP to reduce the dimensionality and HDBSCAN to group similar documents. Finally, c-TF-IDF was used to identify the terms that best represented each cluster, and the researchers interpreted the meaning of the resulting topics.

Another distinctive feature was that the number of topics was not determined arbitrarily. The researchers conducted 5,000 modeling iterations using randomly selected combinations of UMAP and HDBSCAN hyperparameters.

Each solution was evaluated according to the DBCV index, which measures cohesion within clusters and separation between clusters, as well as the total number of topics and the number of publications left unassigned.

The analysis initially generated 54 topics, covering approximately 85 percent of the abstracts. The remaining 15 percent were not forced into existing categories because doing so could have introduced classification errors.

The researchers then merged eight topics with similar content and excluded eight topics that could not be meaningfully interpreted. This process produced a final set of 38 topics.

These 38 topics covered approximately 74 percent of the original dataset. The objective was not to classify every publication without exception, but to produce a set of coherent and interpretable research topics.

3,645 Mentions of LLMs and 1,471 of ChatGPT: The Strong Text-Centric Bias of Educational AI

The first major result was the overwhelming influence of large language models.

An analysis of how frequently major keywords appeared in article titles and abstracts showed that terms related to “large language model” occurred 3,645 times. “Generative artificial intelligence” appeared 2,177 times, while “ChatGPT” appeared 1,471 times.

Although the name ChatGPT was not included in the original search query, it appeared in the collected abstracts far more frequently than Gemini or Bard, Copilot, Claude, or Perplexity.ai. This suggests that ChatGPT has effectively become the representative research object for generative AI in education.

By contrast, text-to-speech, the most frequently mentioned non-text modality, appeared only 461 times. It was followed by text-to-image, text-to-text, speech-to-text, text-to-video, and image-to-image.

Table 1. Frequency of Major Keywords in Generative AI Education Research

These figures indicate that what is currently described as “generative AI research in education” is, in practice, much closer to research on large language models.

Although generative AI can produce many forms of information, educational research has concentrated largely on interactions in which users enter questions as text and receive text-based responses.

Figure 2. Keyword Frequencies of Different AI Modalities in Education Research

Speech and Images Are Possible, but They Remain at the Margins

The second most frequently researched technology after large language models was text-to-speech.

Text-to-speech reads written information aloud and can support reading comprehension and spelling development. It can also improve accessibility for visually impaired students and learners who experience difficulties with reading.

Speech-to-text can help students who struggle with writing express their ideas orally and convert them into written language. This technology may be especially valuable in special education and learning-support contexts because it expands the ways in which students can participate.

Text-to-image has been used to support design ideation, visualize concepts in structural engineering, and produce facial images for medical education. Image-to-image technology has also been applied to generate realistic visualizations of complex molecular structures.

Text-to-video research has explored the creation of educational content and personalized learning materials.

However, compared with research on large language models, these studies remain limited in number and are often confined to specific applications. Although generative AI can connect multiple senses and forms of representation, educational research often treats each modality as an isolated tool.

The researchers found very few studies in which multiple modalities were integrated to generate instructional materials. The term “multimodal AI” may be widely used, but research combining text, speech, images, and video into a coherent learning experience is still at an early stage.

What 38 Topics Reveal About the Current State of Generative AI in Education

The final 38 topics were organized into 14 broader thematic areas:

• Educational domains and disciplines

• Theoretical and philosophical foundations

• Personalized learning support

• Problem-solving

• Technology acceptance

• Teacher professional development

• Creativity

• Serious games

• Learning tools and content

• Assessment

• Ethics and security

• Academic integrity

• Educational chatbots

• Language learning

This thematic map demonstrates that generative AI is not limited to a single educational activity. Research is taking place across nearly every area of education, including instructional-material development, personalized feedback, problem-solving, assessment, teacher education, creative work, foreign-language learning, game development, and research support.

Figure 3. Fourteen Thematic Areas of Generative AI Education Research Identified Through Topic Modeling

The largest of the 38 topics was “LLMs in medical education,” which included 654 abstracts. The second-largest was “text-to-speech for learning support,” with 546 abstracts.

This was followed by generative AI in computing education with 249 abstracts, writing support and feedback with 160, ethical and security challenges with 139, and educational chatbots with 129.

Medical education research frequently evaluated whether AI could simplify patient-education materials or accurately answer medical questions. Computing education research focused on code generation, feedback on programming assignments, and support for novice problem-solving.

In mathematics and physics education, researchers examined the problem-solving ability of language models, the accuracy of their explanations, and their potential use in simulations. In chemistry education, major topics included prompting strategies and the learning of molecular and chemical concepts.

Generative AI is spreading across education, but research is not progressing at the same rate across all school levels and academic disciplines.

Higher Education Moved Quickly, but K–12 Research Remains Limited

Higher education was the only educational level that emerged as an independent topic in the analysis.

“Generative AI in higher education” formed a distinct research cluster, whereas primary and secondary education did not emerge as large, independent topic areas.

This indicates that generative AI research has expanded rapidly in universities while remaining relatively limited in K–12 education. Studies involving primary and secondary students exist, but they have not yet accumulated sufficiently to form a clearly defined research cluster.

The difference is not simply a matter of publication volume. In universities, it is relatively easy to observe students using generative AI to write, generate code, and locate research materials.

In primary and secondary education, however, researchers must consider students’ developmental stages, curriculum requirements, privacy protection, teacher supervision, parental consent, and the reliability of AI-generated outputs.

Using generative AI in K–12 education therefore requires more than scaling down higher-education models. Instructional designs and safety measures appropriate to each school level must be developed.

Personalized Learning Is Prominent, but the Meaning of “Personalization” Remains Narrow

Personalized learning support occupied a relatively large portion of the research landscape. It included text-to-speech, sentiment and emotion analysis, writing support, feedback generation, and intelligent tutoring systems.

Instead of giving every student the same answer to the same question, generative AI can adjust explanations according to students’ responses, rephrase incorrect answers, and provide feedback aligned with individual learning speeds.

Its ability to interact through natural language offers greater flexibility than many conventional adaptive learning systems.

However, personalization should not simply mean generating different sentences for different students. If generative responses do not account for learners’ prior knowledge, misconceptions, goals, motivation, and cognitive load, they may repeat essentially the same educational content in slightly different forms.

Multimodal AI could expand the meaning of personalized learning. It could provide spoken explanations to students who struggle with reading, offer images or simulations to learners who have difficulty understanding abstract concepts, and allow students with language-expression difficulties to respond through speech, gestures, or drawings.

The central purpose of personalization is not merely to modify the wording of an answer. It is to enable students to select the forms of information through which they can most effectively understand and express ideas.

Does Generative AI Support Creativity or Replace It?

“Creativity, art, and design” and “text-to-image for visual design and creativity” also emerged as independent research topics.

Text-to-image technology can help students explore ideas, compare design alternatives, and rapidly visualize possibilities that may previously have been difficult to imagine.

In fashion and textile design education, researchers examined the use of generative AI for idea development and sustainable design exploration. Art-education studies investigated how students and AI systems could co-create digital artworks.

However, as the speed of image generation increases, the boundary between creation and selection becomes less clear.

Even when students do not draw an image themselves, they may design prompts, compare outputs, select results, and determine how an image should be revised. Creativity may therefore shift from manual production toward exploring possibilities, comparing alternatives, and making meaningful judgments.

On the other hand, when students submit AI-generated images without modification, their learning process becomes difficult to observe. It also becomes unclear whether AI supported their creativity or replaced their thinking.

Creative education using multimodal generative AI should therefore evaluate more than the final product. Changes in prompts, comparisons among generated outputs, selection criteria, revision processes, and ethical judgments should also be examined.

The way students collaborate with AI may become a more important assessment target than the generated artwork itself.

The Tension Between a Learning-Support Tool and a Tool for Misconduct

Assessment, academic integrity, ethics, and security were closely connected within the research landscape.

Generative AI can evaluate students’ written responses, automatically generate multiple-choice questions, and provide personalized feedback. These capabilities may reduce teachers’ assessment workload and enable immediate feedback.

At the same time, generative AI can produce assignments, essays, and programming code on behalf of students. Research on detecting AI-generated text and code therefore emerged as independent topics.

Institutional AI policies, plagiarism, the originality of student work, and the authenticity of submitted assignments have become major research concerns.

The same technology can act as a tutor that supports learning or as a tool that replaces student performance.

A student who writes a draft and revises it using AI feedback is engaged in a different learning process from a student who submits a complete AI-generated response. Although the final products may appear similar, their educational meanings are different.

The key issue is not simply whether AI was used, but the extent to which the student participated in knowledge construction and judgment.

Assessment in the era of generative AI should therefore move from a product-centered approach toward a process-centered model. It should consider what questions students asked, how they evaluated AI responses, and what they revised, accepted, or rejected.

Multimodal AI Designed Specifically for Education Is Needed

The researchers argue that education must move beyond simply adopting general-purpose generative AI tools.

Models developed for general users do not necessarily provide responses that align with curricula, learning objectives, or learners’ developmental levels.

Educational AI must do more than generate correct answers quickly. It should diagnose misconceptions, give students time to think, ask questions aligned with learning goals, and clearly identify responses that lack sufficient evidence.

The paper highlights the need for models such as EdGPTs that are intentionally trained to meet educational requirements.

When such models are combined with multimodal functions involving speech, images, and video, students may be able to compare and manipulate different representations rather than passively receiving a single answer.

For example, instead of explaining a mathematical problem only through text, an AI system could provide equations, graphs, changing geometric figures, and spoken explanations.

In science education, it could present experimental procedures through video and simulations, understand students’ spoken questions, and visualize relevant parts of the process again.

The educational value of multimodal AI does not lie simply in displaying more media. Its value lies in helping learners understand relationships among different representations and move meaningfully between them.

Teachers Need “Pedagogical AI Imagination”

The researchers recommend that teachers directly explore and experiment with a broad range of AI technologies, including text-to-speech, speech-to-text, and text-to-image.

The paper describes this as the development of “pedagogical AI imagination.” This means going beyond learning how to use new functions and imagining how a particular AI technology could transform the design of an existing lesson.

Text-to-speech, for instance, can be used simply to read a textbook aloud. It can also allow students to listen to their own writing and revise the flow of their sentences.

Speech-to-text can serve not only as a transcription tool but also as a learning tool that connects speaking and writing.

Text-to-image can be used not merely to produce decorative images, but to reveal students’ conceptual understanding visually and allow different interpretations to be compared.

The educational effectiveness of AI is not embedded in the technology itself. It depends on the order in which it is used, the choices students are given, and the ways in which generated outputs are reviewed.

Teacher education should therefore extend beyond prompt-writing techniques and instructions for using AI services. It should develop teachers’ ability to design AI-supported learning activities, examine the reliability and bias of generated outputs, and protect students’ independent thinking.

Analyzing 4,175 Publications Does Not Reveal Every Aspect of Educational AI

Topic modeling of a large body of literature can reveal broad trends in educational AI research. However, it cannot capture every detail of instructional design or determine the actual learning effects reported in individual studies.

The researchers acknowledge that topic modeling provides a high-level research landscape but may conceal small or emerging topics.

Some publications may also have been omitted because the search query did not include the names of specific AI services.

The lack of standardized terminology presents another challenge. “Copilot” may refer to a general AI assistant, Microsoft Copilot, or GitHub Copilot. Similarly, “perplexity” may refer either to an AI search service or to a statistical measure used to evaluate language models.

The initial search query also did not include “teacher” and “teaching” as independent primary keywords.

Although teacher education and professional development appeared among the resulting topics, the overall study was designed primarily around students and schools. Future reviews should examine how teachers understand and integrate AI into their teaching practices.

Additional modalities, such as text-to-sign-language conversion, were also not fully represented in the search query.

Because the range of multimodal AI technologies continues to expand, research classifications and search terminology must be continuously updated.

The Possibilities of Educational AI Become Visible Only When We Move Beyond Text

The landscape revealed by this study is paradoxical.

Generative AI has advanced rapidly enough to produce images, speech, video, and music, yet educational research remains concentrated on text-based conversations and writing generation.

Large language models clearly offer powerful possibilities for writing, feedback, problem-solving, programming, and assessment. However, students do not learn through text alone.

They develop understanding by seeing, listening, speaking, moving, and connecting multiple forms of representation.

Researching multimodal AI in education therefore means more than adding image generators or speech-synthesis tools. It means identifying which forms of information learners need and organizing text, speech, images, video, and simulations into a coherent learning process.

The next stage of generative AI in education may not depend on introducing an even larger language model.

It may instead depend on connecting existing AI modalities to learners’ characteristics and the nature of each discipline, while encouraging students to compare, verify, and reconstruct generated information rather than consume it passively.

When educational AI moves beyond the chatbox, it can evolve from a tool that writes answers on behalf of students into a learning environment that expands how they construct and express knowledge.