Artificial Intelligence in Education (AIEd) Symposium 2026

In mathematics, computer science and engineering, AI now produces outputs that look like competence. It returns correct solutions, working code and plausible proofs in seconds. This is where the central risk of AI for education is sharpest. The visible output, the performance, is easily mistaken for the invisible capability, the learning. A correct answer is not evidence of understanding, and the gap between the two is widest in the disciplines where the answer is most checkable. These are also the disciplines where students are adopting AI fastest.

This keynote suggests that the principal risk of AI is misjudged. The danger is not that machines become too capable. It is that we undervalue ourselves and allow performance gains to stand in for learning gains. For STEM universities, the opportunity is to design education that develops what AI cannot: accurate judgement of one’s own knowledge, the disposition to question a confident output, and the metacognitive capacity to know when one has understood something and when one has not.

Prof Luckin sets out what this means for the curriculum, assessment and teaching of STEM disciplines, and why the institutions that learn fast and act more slowly will be the ones best placed to prepare their graduates for an AI world: graduates whose value lies in the judgement, reasoning and self-knowledge that no machine supplies.

As AI continues to evolve at an unprecedented pace, the educational landscape is being transformed in ways that challenge traditional learning paradigms. This keynote talk will address the intersection of rapid AI advancements and the nuanced, reflective nature of human learning. It will discuss the distinction between AI for learning and AI for performance, urging AIED designers to prioritise genuine “slower” effortful learning processes over “faster” learning outcomes and solutions.

From a learning science perspective, Prof Chen Wenli will examine how AI-augmented learning environments can be designed not just as tools for quick answers, but as cognitive partners that enhance human agency, foster self-regulation, critical thinking, and metacognitive skills. Drawing on her empirical research, Prof Chen will share human-centric AIED designs to enhance their cognitive and regulatory capacities, rather than undermining them. This talk advocates a shift in focus from efficiency (faster) to meaningful learning (slower), highlighting the importance of human learners’ deep cognitive engagement and agency in human-AI collaboration for learning.

AI-personalised learning has become routine in STEM, yet social-emotional learning (SEL) — arguably a harder and higher-stakes domain — remains almost untouched by AI in terms of learning. This is not an accident but due rather to the difficulty: SEL outcomes are latent, delayed, socially entangled, and costly to measure, and the cost of error is potentially high. We argue that these properties make SEL a revealing testbed for the science of learning, because it forces the questions — measurement, causal identification, and safety — that STEM allows us to skip. Our premise is that AI personalisation is not merely a better delivery mechanism but can also be a causal instrument in analysing learning. By manipulating delivery parameters for a single learner, we can use an SEL intervention as an experiment to answer whether, through which parameters, and by how much AI personalisation can improve SEL. In this talk we cover several technical  topics and open problems in AI for social-emotional learning.

We present CodeHinter, an AI-assisted debugging tool integrated into Visual Studio Code that helps novice programmers debug semantic errors while promoting productive struggle. Rather than providing complete solutions, CodeHinter combines fault localization, interactive hints and quizzes, print-statement suggestions, and memory graph visualization to guide learners through the debugging process. In a pilot study with undergraduate students, the tool was found to be effective, intuitive, and easier to use than its earlier version. Participants especially valued its error-localization and end-to-end testing features. Our findings suggest that AI-based debugging tools should support active problem solving and be personalized to learners’ needs.

Generative AI can rapidly produce concepts convincing, yet many remain unconstrained by physical reality. Consequently, STEM education faces a new challenge: Helping students distinguish physically valid solutions from merely plausible outputs.

This presentation shares a Design•AI (D•AI) teaching approach implemented in Structures & Materials, a core course for Engineering Product Development undergraduates at the Singapore University of Technology and Design (SUTD). Rather than positioning AI as an answer generator, the learning experience places physics at the center of the design process: Students first create their own designs, use them as the basis for AI-assisted exploration, evaluate competing design concepts through physics-based simulations, and finally validate their decisions through fabrication and destructive testing.

Through this approach, we argue that the primary educational outcome for AI in Education is not to produce more proficient AI users, but to develop designers who know when AI can be trusted, when it should be challenged, and how disciplinary knowledge remains the foundation for responsible AI use. The approach illustrates one practical implementation of SUTD’s D•AI principles and offers a transferable pedagogical framework for integrating AI into technically rigorous STEM education while preserving disciplinary rigor and human judgment.