Last week, I had the privilege of presenting our paper, Generative AI for Multiple Choice STEM Assessments, at the EDULEARN Conference in Palma de Mallorca, Spain.
Each year, EDULEARN brings together more than 800 educators, researchers, and education technology leaders from around the world to share ideas, challenge assumptions, and explore how innovation can meaningfully improve teaching and learning. Set against the backdrop of one of Europe’s most beautiful coastal cities, the conference provided an inspiring setting. But what stood out most wasn’t the venue. It was the people.
One theme emerged repeatedly throughout the week: technology should never be adopted for its own sake.
Across sessions and conversations, there was a shared commitment to using technology thoughtfully; solving real educational challenges, supporting educators, and ultimately improving outcomes for students. That philosophy has long guided our work at DigitalEd, making EDULEARN an especially rewarding community to engage with.
Our paper explored a question that sits at the intersection of artificial intelligence, assessment, and mathematics education:
Can the same characteristics that limit AI’s mathematical reasoning become useful for designing better assessments?
Large language models have well-documented limitations when it comes to symbolic reasoning and multi-step mathematical problem solving. They often produce solutions that appear convincing, yet contain procedural mistakes or flawed reasoning. Traditionally, these shortcomings are viewed as obstacles to using AI in STEM education.
We approached the problem from a different perspective.
Rather than asking AI to generate perfect mathematical solutions, we investigated whether those same reasoning errors could be leveraged to create something assessment authors often struggle to produce: high-quality distractors for multiple-choice questions.
Effective distractors are far more than simply incorrect answers. They must reflect authentic misconceptions; mistakes that real students are likely to make when solving a problem. Creating these plausible alternatives is widely recognized as one of the most difficult and time-consuming aspects of assessment authoring, particularly in STEM disciplines.
Our work demonstrated that, when paired with computational validation, generative AI can become a valuable assistant in this process.
Within the Möbius platform, AI generates candidate distractors while the underlying mathematics engine offers the ability to validate their correctness, ensuring that only one answer is mathematically correct and that the remaining options remain distinct, plausible, and instructionally meaningful. In other words, AI provides the creative possibilities, while deterministic computation preserves academic rigour.
The response from attendees reinforced something we’ve believed for some time: educators are not looking for AI that replaces expertise. They’re looking for AI that amplifies it.
Throughout the conference, discussions repeatedly returned to practical questions of trust, validation, transparency, and instructional value. These are precisely the questions we sought to address in our research. The future of AI in education will not be determined by how much content it can generate, but by how thoughtfully that content is integrated into sound pedagogical practice.
That perspective closely mirrors our mission at DigitalEd.
One of my favorite moments from the presentation came during the discussion afterward, when several attendees reflected on a simple but powerful idea:
Perhaps the value of AI in STEM assessment isn’t in solving mathematical problems perfectly, but in helping educators create better ways to evaluate how students think.
For me, that observation extends far beyond this single research project.
In Season 4 of The EdTech Curio Cabinet, I explore a central question: What remains fundamentally human when intelligent machines become increasingly capable?
As AI begins to explain concepts, generate content, and even participate in assessment, the real challenge is no longer whether machines can produce answers. It is whether they can help us better understand the learning that leads to those answers.
In many ways, our research reflects that same shift in perspective. Rather than asking AI to replace human expertise, we asked how it might amplify it. Rather than expecting perfect mathematical reasoning from a language model, we explored how its limitations could help educators uncover authentic student misconceptions and, ultimately, create richer opportunities for learning.
Perhaps that is the larger lesson emerging from both this research and the conversations at EDULEARN26. The future of educational technology will not be defined by how intelligently machines perform educational tasks, but by how intentionally we use those capabilities to cultivate curiosity, reasoning, and understanding in learners.
Because, in the end, education has never been about producing answers.
It’s about developing people who know how to think.