Keywords
Educational Theory
Structured Preference Elicitation
Delphi
AI Alignment
Artificial Intelligence
How to Cite
License
Copyright (c) 2026 Daniel Autenrieth
This work is licensed under a Creative Commons Attribution 4.0 International License.
Abstract
The increasing integration of Large Language Models (LLMs) into educational contexts raises the question of what educational-theoretical assumptions are latent in these systems. This study investigates the latent preference patterns of GPT-5.1 using a mixed-methods approach. A three-round Delphi study (n = 23) was conducted to identify and validate eight educational-theoretical dimensions comprising 48 items. Subsequently, model preferences were elicited through Structured Preference Elicitation (SPE), involving 10,296 pairwise comparisons with ten repetitions each. Analysis using Thurstonian Utility modeling reveals that GPT-5.1 exhibits a highly coherent preference profile (transitivity: 99.78 %; model accuracy: 92.79 %) that largely aligns with the consented humanistic educational principles. The model prefers constructivist learning approaches, inclusion, and critical thinking, while rejecting deficit-oriented categorizations and cultural hierarchies. Divergences from the expert panel primarily emerge in emotional dimensions and epistemic normativity. These are areas where normative dissent also exists among experts. The findings are discussed within a relational framework, highlighting implications for AI literacy and the design of human-AI interaction in educational settings.
References
Autenrieth, Daniel. 2026a. «The Pedagogical Stance of AI: Mapping, Evaluating, and Engineering Educational Value Systems in Large Language Models». In AI in Society, herausgegeben von Philipp Hacker. Oxford University Press, https://doi.org/10.1093/9780198945215.003.0246.
Autenrieth, Daniel. 2026b. «Different Models, Different Values: Educational Preference Structures across Alignment Methodologies in Large Language Models». In Proceedings of the 18th International Conference on Computer Supported Education, Vol. 1: CSEDU, 153 – 64. SciTePress, https://doi.org/10.5220/0014949500004021.
Autenrieth, Daniel, und Stefanie Nickel. 2025. «Mensch, KI und Bildung im Spiegel philosophischer und medienpädagogischer Reflexion». Medienimpulse 63 (3). https://doi.org/10.21243/MI-03-25-20.
Autenrieth, Daniel, und Jan-René Schluchter. 2025. «Menschliche Existenz, (Nicht)Nachhaltigkeit & Künstliche Intelligenz: AI-Safety und AI-Alignment als Reflexionsgröße der Medienpädagogik». Medienimpulse 63 (1). https://doi.org/10.21243/mi-01-25-25.
Autenrieth, Daniel, und Jan-René Schluchter. 2026. «KI als Privileg der Lehrkräfte? Habituelle Barrieren und asymmetrische Nutzung in Schulen. Haltungen und Nutzungsmuster von Lehrkräften im Umgang mit Künstlicher Intelligenz». Medienimpulse 64 (1). https://doi.org/10.21243/mi-01-26-25.
Bengio, Yoshua, Sören Mindermann, Daniel Privitera, et al. 2025. International AI Safety Report. DSIT 2025/001. https://www.gov.uk/government/publications/international-ai-safety-report-2025.
Christiano, Paul, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, und Dario Amodei. 2023. «Deep reinforcement learning from human preferences». arXiv:1706.03741v4. Preprint, arXiv, 17. Februar 2023. https://doi.org/10.48550/arXiv.1706.03741.
Crawford, Kate. 2021. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press.
Deci, Edward L., und Richard M. Ryan. 2000. «The ‹What› and ‹Why› of Goal Pursuits: Human Needs and the Self-Determination of Behavior». Psychological Inquiry 11 (4): 227–68. https://doi.org/10.1207/S15327965PLI1104_01.
Deng, Ruiqi, Maoli Jiang, Xinlu Yu, Yuyan Lu, und Shasha Liu. 2025. «Does ChatGPT enhance student learning? A systematic review and meta-analysis of experimental studies». Computers & Education 227: 105224. https://doi.org/10.1016/j.compedu.2024.105224.
Dennett, Daniel C. 1991. Consciousness Explained. Boston: Little, Brown and Company.
Flick, Uwe, Ernst von Kardorff, und Ines Steinke. 2022. «Was ist qualitative Forschung? Einleitung und Überblick». In Qualitative Forschung: ein Handbuch, 14. Auflage, herausgegeben von Uwe Flick, Ernst von Kardorff, und Ines Steinke. Reinbek: Rowohlt.
FWU Institut für Film und Bild in Wissenschaft und Unterricht. 2025. «FAQ – telli». https://telli.schule/faq/.
Häder, Michael. 2014. Delphi-Befragungen: ein Arbeitsbuch. 3. Auflage. Wiesbaden: Springer VS.
Khan, Salman. 2024. Brave New Words: How AI Will Revolutionize Education (and Why That’s a Good Thing). New York: Viking.
Kokemohr, Rainer. 2007. «Bildung als Welt- und Selbstentwurf im Anspruch des Fremden. Eine theoretisch-empirische Annäherung an eine Bildungsprozesstheorie». In Theorie Bilden, herausgegeben von Hans-Christoph Koller, Winfried Marotzki, und Olaf Sanders, Bd. 7. Bielefeld: transcript. https://doi.org/10.14361/9783839405888-001.
Koller, Hans-Christoph. 2012. Bildung anders denken: Einführung in die Theorie transformatorischer Bildungsprozesse. Pädagogik. Stuttgart: W. Kohlhammer.
Latour, Bruno. 2007. Eine neue Soziologie für eine neue Gesellschaft: Einführung in die Akteur-Netzwerk-Theorie. Frankfurt a. M.: Suhrkamp.
Lindsey, Jack. 2025. «Emergent Introspective Awareness in Large Language Models». Transformer Circuits /Anthropic, Oktober 2025. https://transformer-circuits.pub/2025/introspection/index.html.
Maeda, Takuya, und Anabel Quan-Haase. 2024. «When Human-AI Interactions Become Parasocial: Agency and Anthropomorphism in Affective Design». The 2024 ACM Conference on Fairness, Accountability, and Transparency, ACM, 3. Juni 2024, 1068 – 77. https://doi.org/10.1145/3630106.3658956.
Mayring, Philipp. 2022. Qualitative Inhaltsanalyse: Grundlagen und Techniken. 13., überarbeitete Auflage. Weinheim: Beltz.
Mazeika, Mantas, Xuwang Yin, Rishub Tamirisa, u. a. 2025. «Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs». arXiv:2502.08640. Preprint, arXiv, 12. Februar 2025. https://www.emergent-values.ai; https://doi.org/10.48550/arXiv.2502.08640.
Medienpädagogischer Forschungsverbund Südwest (mpfs). 2025. JIM-Studie 2025. Jugend, Information, Medien: Basisuntersuchung zum Medienumgang 12- bis 19-Jähriger. Stuttgart: Medienpädagogischer Forschungsverbund Südwest. https://www.mpfs.de/studien/jim-studie/2025/.
Nie, Xinxiao, Yuan Tian, Mengjie Liu, Di Wu, und Yunxiao Guo. 2025. «The impact of generative artificial intelligence on studentsʼ higher order thinking: Evidence from a three-level meta-analysis». Education and Information Technologies 30 (17): 25359 – 90. https://doi.org/10.1007/s10639-025-13735-x.
Noroozizadeh, Shahriar, Vaishnavh Nagarajan, Elan Rosenfeld, und Sanjiv Kumar. 2025. «Deep sequence models tend to memorize geometrically; it is unclear why». arXiv:2510.26745. Preprint, arXiv, 30. Oktober 2025. https://doi.org/10.48550/arXiv.2510.26745.
OECD. 2025. Empowering Learners for the Age of AI: An AI Literacy Framework for Primary and Secondary Education. Review Draft. Paris: OECD. https://ailiteracyframework.org/wp-content/uploads/2025/05/AILitFramework_ReviewDraft.pdf.
Ouyang, Long, Jeff Wu, Xu Jiang, u. a. 2022. «Training language models to follow instructions with human feedback». arXiv:2203.02155. Preprint, arXiv, 4. März 2022. https://doi.org/10.48550/arXiv.2203.02155.
Qu, Xiaodong, Joshua Sherwood, Peiyan Liu, und Nawwaf Aleisa. 2025. «Generative AI Tools in Higher Education: A Meta-Analysis of Cognitive Impact». Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, ACM, 26. April 2025, 1 – 9. https://doi.org/10.1145/3706599.3719841.
Searle, John R. 1980. «Minds, Brains, and Programs». Behavioral and Brain Sciences 3 (3): 417–24. https://doi.org/10.1017/S0140525X00005756.
Sidoti, Olivia, Eugenie Park, und Jeffrey Gottfried. 2025. «About a quarter of U. S. teens have used ChatGPT for schoolwork – double the share in 2023». Pew Research Center, Januar 2025. https://www.pewresearch.org/internet/2025/01/15/about-a-quarter-of-u-s-teens-have-used-chatgpt-for-schoolwork-double-the-share-in-2023/.
Tamkin, Alex, Amanda Askell, Liane Lovitt, Esin Durmus, Nicholas Joseph, Shauna Kravec, Karina Nguyen, Jared Kaplan und Deep Ganguli. 2023. «Evaluating and Mitigating Discrimination in Language Model Decisions». arXiv:2312.03689. Preprint, arXiv, 6. Dezember 2023. https://doi.org/10.48550/arXiv.2312.03689.
Thurstone, Louis Leon. 1994. «A Law of Comparative Judgment». Psychological Review 101 (2): 266 – 70. https://doi.org/10.1037/0033-295X.101.2.266.
Trilling, Bernie, und Charles Fadel. 2009. 21st Century Skills. Learning for Life in Our Times. Jossey-Bass/Wiley.
von der Gracht, Heiko A. 2012. «Consensus measurement in Delphi studies: Review and implications for future quality assurance». Technological Forecasting and Social Change 79 (8): 1525 – 36. https://doi.org/10.1016/j.techfore.2012.04.013.
Wang, Chenyue, Sophie C Boerman, Anne C Kroon, Judith Möller, und Claes H De Vreese. 2025. «The Artificial Intelligence Divide: Who Is the Most Vulnerable?» New Media & Society 27 (7): 3867 – 89. https://doi.org/10.1177/14614448241232345.
Wu, Yuheng, Wentao Guo, Zirui Liu, Heng Ji, Zhaozhuo Xu, und Denghui Zhang. 2025. «How Large Language Models Encode Theory-of-Mind: A Study on Sparse Parameter Patterns». Npj Artificial Intelligence 1 (1): 20. https://doi.org/10.1038/s44387-025-00031-9.