The DOI matches its Crossref record; 103 of the 103 references it cites were found in a registry or an index; 13 of the 108 sources its claims cite were checked for support.
On opportunities and challenges of large multimodal foundation models in education
Open accessAbstract
Recently, the option to use large language models as a middleware connecting various AI tools and other large language models led to the development of so-called large multimodal foundation models, which have the power to process spoken text, music, images and videos. In this overview, we explain a new set of opportunities and challenges that arise from the integration of large multimodal foundation models in education.
Overview
CoChat Summary
TL;DR
This overview argues that large multimodal foundation models (LMFMs) such as ChatGPT-4-Turbo and Gemini create new educational opportunities for learners, teachers, researchers, and developers by processing text, speech, images, and video, while also introducing challenges around accuracy, bias, privacy, access, and overreliance. It proposes mitigation strategies emphasizing AI literacy, human oversight, transparent documentation, and equitable access.
Key findings
- LMFMs extend large language models by enabling input and output via written text, spoken text, images, and videos, and can act as middleware connecting different AI tools; examples named include ChatGPT-4-Turbo and Gemini.
- Opportunities span personalized learning, barrier-free interaction, multimedia learning, representational competence training, experimentation and code-generation support, teacher lesson planning and assessment, researcher data analysis and literature review, and developer creation of customized learning applications.
- Challenges include output and source verification, difficulty maintaining an overview of involved AI models, potential segregation due to access costs, ethical issues and inherent biases, dependency and overuse, personal data leaks, plagiarism risks, digital divide, erosion of human relations, reliability of information, and need for AI literacy.
- The paper reports that GPT-4 Turbo’s expanded context length can consider and modify extensive source code, potentially encompassing 128k tokens, described as more than 300 pages of text.
Methods
- Design: narrative overview/perspective that synthesizes prior literature, examples, and arguments about LMFMs in education; it is not an empirical study.
- Data or sample: Not stated in the paper; no dataset, participants, or systematic sampling procedure is reported.
- Main analysis: Not stated in the paper; no formal quantitative or qualitative analysis is described.
Limitations
- The paper does not state limitations of its own overview.
- Evident from the text: it is a non-systematic overview/perspective, so its claims about opportunities and challenges are not empirically tested in this manuscript.
- Evident from the text: the authors note increasingly short innovation cycles in AI, which makes specific model examples and claims potentially time-sensitive.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
9 Valid · 1 Not valid · 66 Not checked · 108 Own results and interpretations
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“ChatGPT, for instance, has been favorably received by students19,20.”
§ On opportunities and challenges of large multimodal foundation models in education > Opportunities > Opportunities for learners
What the sources say
- Supportedcites [19] Romero Rodríguez, J. M. et al. Use of ChatGPT at university as a tool for complex thinking: Students’ perceived usefulness (Universidad de Alicante, 2023). source
The Discussion (p. 10) states that the results 'showed assent and recognition of the potential of this chatbot in the learning process' and that 'Acceptance by university students and evidence about its positive impact on the educational process can lead to educational innovations.' The Conclusions (p. 13) likewise report 'the acceptance of ChatGPT by university students,' driven by experience of use, performance expectancy, hedonic motivation, price value and habit shaping behavioural intention and user behaviour (Abstract, p. 1; Discussion, p. 12). This is a broad background statement, and our copy clearly treats ChatGPT as favourably received by (university) students, so the paraphrase is supported; the study's sample is 400 Spanish university students, which the citing sentence does not specify but does not misrepresent. (LLM, conf=0.85)
- Disputedcites [20] Singh, H., Tayarani-Najaran, M.-H. & Yaqoob, M. Exploring Computer Science Students’ Perception of ChatGPT in Higher Education: A Descriptive and Correlation Study. Educ. Sci. 13, 924 (2023). sourceExperimental — accuracy not yet measured
Our copy of Singh et al. (2023) reports the opposite of a favorable reception: the Abstract (p. 1) states that 'students are skeptical of its positive impacts on learning' and that many are familiar with the tool but do not regularly use it for academic purposes. The Results (p. 9) show a split, 'rather flat' distribution on impact — 38% believing ChatGPT has a negative impact on international students versus 37% a positive one, with 24% seeing neither — and the Conclusions (p. 18) emphasize student concern about misuse and a lack of in-depth skills/knowledge. So this survey does not support 'favorably received by students'; it contradicts it. (LLM, conf=0.85)
Checked by a reasoning model — may be wrong.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“With advanced context understanding, LMFMs can facilitate collaborative learning through peer feedback models.”
§ On opportunities and challenges of large multimodal foundation models in education > Opportunities > Opportunities for learners
What the sources say
- Partly supportedcites [13] Bauer, E. et al. Using natural language processing to support peer-feedback in the age of artificial intelligence: A cross-disciplinary framework and a research agenda. Br. J. Educ. Technol. 54, 1222–1245 (2023). sourceExperimental — accuracy not yet measured
The opening summary (p. 1) says large language models such as ChatGPT and GPT-4 have the potential to transform peer feedback, and the Introduction (p. 3) describes NLP-based adaptive support in digital peer-feedback environments. The paper does not establish the claim specifically for LMFMs or attribute that potential to their advanced context understanding. (LLM, conf=0.84)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The authors also uncovered ChatGPT’s capacity to identify and address ethical, governance, and socio-technical considerations within this context66.”
§ On opportunities and challenges of large multimodal foundation models in education > Opportunities > Opportunities for developers of educational applications
What the sources say
- Partly supportedcites [66] Ahmad, A. et al. Towards human-bot collaborative software architecting with chatgpt. In Proceedings of 27th International Conference on Evaluation and Assessment in Software Engineering, 279–285 (ACM, 2023). sourceExperimental — accuracy not yet measured
The Introduction and Conclusions identify ethics, governance, and socio-technical issues as considerations in collaborative architecting. Section 3.2 says ChatGPT can outline constraints when queried, including compliance with data-security policies. But the Ethics and Intellectual Property section warns that ChatGPT-generated code may create privacy or regulatory problems and assigns the architect responsibility for addressing them; the paper does not establish ChatGPT’s capacity to identify and address all these considerations. (LLM, conf=0.85)
plainSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Verification of sources: Another challenge arises from the difficulty of verifying the sources from which information is retrieved69.”
§ On opportunities and challenges of large multimodal foundation models in education > Challenges > Challenges for learners
What the sources say
- Partly supportedcites [69] Krupp, L. et al. Unreflected acceptance-investigating the negative consequences of chatgpt-assisted problem solving in physics education. arXiv, https://arxiv.org/abs/2309.03087 (2023). sourceExperimental — accuracy not yet measured
The Discussion (p. 7) reports that students had difficulty determining whether ChatGPT’s answers were correct, supporting a challenge in verifying retrieved information. It does not establish that they had difficulty verifying the sources from which that information came; the study assessed answer correctness, not source provenance. (LLM, conf=0.78)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The ability to upload files with exercises and solutions ensures the precision of the GPT’s responses, a critical feature in educational environments15.”
§ On opportunities and challenges of large multimodal foundation models in education > Opportunities > Opportunities for developers of educational applications
What the sources say
- Partly supportedcites [15] Steinert, S., Avila, K. E., Ruzika, S., Kuhn, J. & Küchemann, S. Harnessing large language models to develop research-based learning assistants for formative feedback. Smart Learn. Environ 11, 62 (2024). sourceExperimental — accuracy not yet measured
The Functions and features of LEAP section (p. 7) says teachers enter a task question and its correct answer; the Discussion (p. 12) says providing the correct solution strongly reduces erroneous outputs and hallucinations. The paper describes entering these in fields, not uploading files, and says errors are reduced rather than that response precision is ensured. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
60 more claims · 3 with their cited sources checked
Create a free account to see all 184 claims and to verify the 95 references still open.
- It may state two separate claims as one. 61
- Probably not a substantial claim. 90
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“By contrast, GenAI tools such as large language models (LLMs)-although likely less precise in instructional alignment and associated with challenges such as overreliance, reliability, and ethical concerns (Han et al. 2025a, b )-are more interactive and flexible in adapting to learners' needs through flexible dialogue without requiring extensive pre -modeling and could be used to support SRL by”
From Enhancing School Students' Self-Regulated Learning through Generative AI Support: A Randomized Controlled Trial (2026) · cites this paper as Küchemann et al., 2025
Experimental — accuracy not yet measured
Citations
Cites
High-Resolution Image Synthesis with Latent Diffusion Models
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01042 · Cited by 20,302 · cited 1×
“Overcoming these challenges requires not only innovative approaches to data collection and sharing in the educational domain, but also the employment of methods for data augmentation and generation 107 – 109”
ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education
Kasneci, E., Seßler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., Gasser, U., Groh, G., Günnemann, S., Hüllermeier, E., et al. (2023) · doi:10.35542/osf.io/5er8f · Cited by 7,552
Self-citationChatGPT for good? On opportunities and challenges of large language models for education
Kasneci, E., Sessler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., Gasser, U., Groh, G., Günnemann, S., Hüllermeier, E., et al. (2023) · Learning and Individual Differences · doi:10.1016/j.lindif.2023.102274 · Cited by 7,326 · cited 2×
“Yet,theissueofthedigitaldivideandacces-sibilityremainsaconcern,asnotalleducationalinstitutionsmayhavethe resourcestoimplementandmaintainsuchadvancedtechnology effectively 1 .”
Self-citationA Survey on Bias and Fairness in Machine Learning
Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., & Galstyan, A. (2022) · ACM Computing Surveys · doi:10.1145/3457607 · Cited by 5,812
The Black Box Society: The Secret Algorithms That Control Money and Information
Pasquale, F. (2015) · doi:10.4159/harvard.9780674736061 · Cited by 3,796
How the machine “thinks”: Understanding opacity in machine learning algorithms
Burrell, J. (2016) · Big Data & Society · doi:10.1177/2053951715622512 · Cited by 3,196
The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring
BLOOM, B. S. (1984) · Educational Researcher · doi:10.3102/0013189x013006004 · Cited by 2,734
Ray, P. P. (2023) · Internet of Things and Cyber-Physical Systems · doi:10.1016/j.iotcps.2023.04.003 · Cited by 2,463
VanLEHN, K. (2011) · Educational Psychologist · doi:10.1080/00461520.2011.611369 · Cited by 2,323 · cited 1×
“Since several years, intelligent tutoring systems (ITS) have been developed and effectively integrated into learning environments to provide enriched personalized learning experiences, for instance, by offering tailored feedback and guidance 60,61 .”
Conceptualizing AI literacy: An exploratory review
Ng, D. T. K., Leung, J. K. L., Chu, S. K. W., & Qiao, M. S. (2021) · Computers and Education: Artificial Intelligence · doi:10.1016/j.caeai.2021.100041 · Cited by 2,205 · cited 1×
“(cid:129) Curriculumdesignandeducationalpolicies: Educationalinstitutions face the challenge of keeping pace with the rapid changes in LLM capabilities, necessitating constant adaptation of curricula and educational policies to ensure relevance and effectiveness 92,93 .”
Research works printed but not linked: 9
- Fauzi, F., Tuhuteru, L., Sampe, F., Ausat, A. M. A. & Hatta, H. R. Analysing the role of ChatGPT in improving student productivity in higher education. J. Educ. 5, 14886–14891 (2023).
- Tavangarian, D. et al. Is e-Learning the Solution for Individual Learning? Electron. J. E learn. 2, 265–272 (2004).
- Hutson, J. & Robertson, B. Exploring the Educational Potential of AI Generative Art in 3D Design Fundamentals: A Case Study on Prompt Engineering and Creative Workflows. Glob. J. Hum.-Soc. Sci. 23, 485 (2023).
- Meske, C. & Bunde, E. Design principles for user interfaces in ai-based decision support systems: The case of explainable hate speech detection. Inf. Syst. Front. 25, 743–773 (2023).
- Goodman, B. & Flaxman, S. European union regulations on algorithmic decision-making and a “right to explanation”. AI Mag. 38, 50–57 (2017).
- Anderson, J., Rainie, L. & Luchsinger, A. Artificial intelligence and the future of humans. Pew Res. Center 10, 12 (2018).
- Trust, T., Whalen, J. & Mouza, C. Chatgpt: Challenges, opportunities, and implications for teacher education. Contemp. Issues Technol. Teach. Educ. 23, 1–23 (2023).
- Leemann, T., Rong, Y., Nguyen, T.-T., Kasneci, E. & Kasneci, G. Caution to the exemplars: On the intriguing effects of example choice on human trust in XAI. In XAI in Action: Past, Present, and Future Applications (NIPS, 2023).
- Borisov, V., Sessler, K., Leemann, T., Pawelczyk, M. & Kasneci, G., Language models are realistic tabular data generators. In: The Eleventh International Conference on Learning Representations (ICLR, 2022).
Cited by
Wang, S., & Zhang, H. (2026) · International Journal of Educational Technology in Higher Education · doi:10.1186/s41239-026-00585-x · Cited by 38
Tay, F. R., Loveless, R., & Ravenel, T. D. (2026) · Dental Research · doi:10.1016/j.dtrs.2025.100005 · Cited by 27
Teacher training in the age of AI: impact on AI literacy and teachers’ attitudes
Lademann, J., Henze, J., Honke, N., Wollny, C., & Becker, S. (2026) · Frontiers in Education · doi:10.3389/feduc.2025.1671306 · Cited by 62
Creating a customisable Socratic AI physics tutor
Tufino, E., & Gregorcic, B. (2025) · Physics Education · doi:10.1088/1361-6552/ae0d23 · Cited by 16
The future of personalized learning with AI
Plass, J. L., & Froehlich, F. (2025) · Learning and Individual Differences · doi:10.1016/j.lindif.2025.102813 · Cited by 18
Metadata
- Authors
- Küchemann, S., Avila, K. E., Dinc, Y., Hortmann, C., Revenga, N., Ruf, V., Stausberg, N., Steinert, S., Fischer, F., Fischer, M., Kasneci, E., Kasneci, G., Kuhr, T., Kutyniok, G., Malone, S., Sailer, M., Schmidt, A., Stadler, M., Weller, J., & Kuhn, J.
- Affiliations
- Ludwig-Maximilians-Universität München (DE)
- Institut für Didaktik und Ausbildungsforschung in der Medizin (DE)
- Technical University of Munich (DE)
- Deutsches Zentrum für Luft- und Raumfahrt e. V. (DLR) (DE)
- Saarland University (DE)
- University of Augsburg (DE)
- Corresponding author
- Küchemann, Stefan
- ORCID
- Venue
- npj Science of Learning
- Published
- 2025-02-26
- Volume
- 10
- Issue
- 1
- Pages
- 11-11
- ISSN
- 2056-7936 (ISSN-L)
- PMC
- PMC11861286
- OpenAlex
- W4407928447
- Semantic Scholar
- b4683e8d44296a30f15f842aafbf283342d40f7b
- Cited by
- 47
- References
- 76
- Citations by year
- 2026: 32 · 2025: 15
- FWCI
- 4.92
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- npj Science of Learning (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- Open access LMU (Ludwid Maxmilian's Universitat Munchen) (submittedVersion)
- OPUS (Augsburg University) (submittedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
- mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) (submittedVersion, other-oa)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
- Social Sciences › Education › Artificial Intelligence in Education
- Physical Sciences › Computer Science › Artificial Intelligence › Intelligent Tutoring Systems and Adaptive Learning
- Keywords
- large language models, artificial intelligence in education, multimodal AI, educational technology integration, opportunities and challenges
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-04, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-04 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash and gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: On opportunities and challenges of large multimodal foundation models in education. https://app.cochat.ai/papers/10.1038/s41539-025-00301-w
- The paper
- Küchemann, S., Avila, K. E., Dinc, Y., Hortmann, C., Revenga, N., Ruf, V., Stausberg, N., Steinert, S., Fischer, F., Fischer, M., Kasneci, E., Kasneci, G., Kuhr, T., Kutyniok, G., Malone, S., Sailer, M., Schmidt, A., Stadler, M., Weller, J., & Kuhn, J. (2025). On opportunities and challenges of large multimodal foundation models in education. npj Science of Learning. https://doi.org/10.1038/s41539-025-00301-w
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.