Back to The EDiT Journal
The EDiT Spotlight with Adaptemy (Part 1)
Voices from the frontlines of education and technology
EDT&Partners
EdTech
Schools & Universities
.png)
The EDiT Spotlight highlights individuals across the education and knowledge ecosystem who are shaping how learning evolves in practice. In each edition, we feature a conversation with a leader or innovator working at the intersection of education, technology, and the future of knowledge.
In this edition, we feature Adaptemy, the Dublin-based adaptive learning company founded by Conor O'Sullivan. Since 2013, Adaptemy has been building the AI and learning science infrastructure that powers personalized learning products for publishers, national education systems, and language learning providers. We invited four members of the team to share where adaptive learning stands today and where it needs to go.
Arthur Alexion, CEO, leads Adaptemy across engineering, commercial, and product teams. His background in mathematics and two decades of scaling growth businesses shape how he thinks about adaptive AI: accurate prerequisite mapping, deterministic models, and technology that can explain its own decisions.
Dr. Ioana Ghergulescu, Chief AI & Learning Science Officer, built the scientific foundations of Adaptemy's AI Engine and has been working on adaptive learning since 2008. Her position is direct: AI in education should start with understanding how people learn, and every recommendation the system makes should be explainable.
Daniel McCrea, Head of Partnerships, spent roughly eleven years leading adaptive learning implementations before moving into his current role. His conviction is that education has reached an inflection point: educators have always known every learner is unique, but lacked the means to act on that at scale.
Jessica Kennedy White, Learning Sciences Researcher and Learning Design Architect, brings a career spanning teaching, EdTech evaluation, and research on metacognition. Her focus is on how learners develop the capacity to think about their own thinking, and she is clear that AI should enrich the human relationships in education, not stand in for them.
In our conversation, the team returns to the concept of "productive struggle": the space where challenge sparks learning. Find out why.
Our Interview with the Adaptemy Team
1. Adaptive learning has been promising to "personalize education" for over a decade. Where has it fallen short, and what must AI in education do beyond personalization?
Jessica White
Adaptive learning can genuinely deliver on reaching students individually at scale. By identifying misconceptions and crucial misunderstandings, adaptive learning can identify key strengths and support students with self-regulation especially with challenging concepts. When there is a ratio of teachers-students that is 1:30, or in some rural areas in Asia 2-4:100, it’s difficult for teachers to manage such large classes. With adaptive learning it is possible to really support students at scale.
However, similar to the medical field, there often isn’t a silver bullet. Rather an integrated and holistic approach is needed, where adaptive learning is one part of the solution. Recently, there has been a lot more development in different kinds of learning design that is known as ‘learning how to learn’. Metacognition has been more integrated into the learning design of platforms, but also, in recent years, it has been explicitly taught in classrooms/universities/workplaces. The multifaceted concept of metacognition is becoming more recognised as being an essential part of learners’ and teachers’ core development. Furthermore, how metacognition manifests in both learning design as well as different contexts, is important to be designed collaboratively, as it changes, depending on those contexts.
In focusing on smooth personalisation, there is danger of having friction designed out of learning platforms, so that students experience ease over ‘productive struggle’. However, that struggle, not too much, but just enough to be challenged to think, is where active learning occurs.
Often it is in that awareness of why we struggle with a concept, question or problem, where we become most aware, which is metacognition in action. This combination of factors that are activated by a student's experience of ‘productive struggle’ whilst engaging with adaptive learning design, can enable durable learning. This is sometimes missed out if a platform focuses too much on routing learners through personalisation to the point where we don’t give students the time and space to struggle and rather just make it learning efficient for them. In such a case, we can inadvertently take away that safe space to struggle, so that they can engage in their deepest thinking in favour of taking them on a dopamine high.
Designing for ‘productive struggle’ both within the platform, but also encouraging this in their blended learning contexts, can give them more space to be challenged and to grow. Furthermore, by making learners’ thinking visible to themselves, they can become better at self-regulation and their understanding of their affective states. This can also support social learning, by creating a learning design that makes explicit their mistakes and processes. This can be framed as a positive part of ‘learning how to learn’, rather than just binary success and failure. Personalisation is the support, designing for learning, where we can encourage self-aware thinking, combined with human flourishing is the purpose.
Daniel McCrea
Adaptive learning has fallen short in scaling beyond successful pilots. It is relatively easy to build one or two beautiful adaptive courses. The real challenge is making adaptive learning a sustainable capability across an institution's entire catalogue.
Two things commonly prevent this: technology that cannot accommodate different pedagogical approaches, and authoring tools so complex that only a handful of specialists can operate them. When those specialists leave, the capability leaves with them.
Beyond personalisation, AI must deliver value to the entire educational ecosystem: giving editors evidence of which content works, helping institutions continuously improve assessment quality, and giving teachers actionable insights. Automate what technology does best, and give teachers back the human part of teaching.
Ioana Ghergulescu
Adaptive learning has made meaningful progress over the past decade, but there is still significant opportunity to extend its reach and deepen its impact, particularly in making true adaptivity available to every learner, across every subject. Successful large-scale implementations have demonstrated what is possible when education ministries and technology partners work together to align curriculum, content, and technology.
However, achieving true adaptivity requires more than adjusting content or learning pathways based on performance. It requires rich, longitudinal learner models and intelligent learning orchestration that continuously respond to each learner's evolving needs. AI presents an opportunity to advance this further, enabling holistic learner models that encompass cognitive abilities, metacognitive skills, learner characteristics, and individual traits, and using these insights to orchestrate truly personalized learning experiences.
Crucially, explainability must be at the heart of this orchestration, empowering learners to understand their learning journeys and enabling teachers and lecturers to understand the rationale behind AI-driven decisions, trust the recommendations, and act on them effectively.
2. In practice, the line between a student who is struggling productively and one who is genuinely stuck can be hard to see, especially in a digital environment. What does learning science tell us about getting that distinction right?
Jessica White
This differentiation can be understood within the context of how and in what ways learning design in digital environments can enable learners to tackle a challenging question, problem or concept that is within their reach. Learning design can provide scaffolding for learners within this space to help them feel safe enough to ease into the experience of ‘productive struggle’, where they can make mistakes and learn from them.
Learners’ analytics triangulated with student’s reflections and teacher’s observations both on a cognitive and affective level, can help. These representations and feedback can be synthesised to support both learners and their teachers with a clearer picture of when students are stuck and when they are in that transformative ‘productive struggle’ space. This can help to adapt learning resources effectively, so that they are not challenged to the point beyond their threshold.
So, when designing for learning, we not only need to sequence learning design in a way that allows for ‘productive struggle’, but also, increasingly, it is important to know how a learner feels about the level of difficulty they are experiencing within the platform. This combination helps learning designers to understand where they are stuck and need more scaffolding or resources, or when they have enough support and we can create that supportive space for learners to solve problems and understand concepts themselves.
Ioana Ghergulescu
Learning science tells us that productive struggle is essential for learning, but the challenge is knowing when that struggle is supporting growth and when it is turning into frustration or disengagement. The key is to become much more accurate at identifying that boundary. This requires a rich and accurate learner model that understands what a student knows and does not yet know, a reliable estimation of content difficulty relative to the learner's abilities, and an understanding of their affective state, motivation, and learning flow.
With these insights, intelligent learning orchestration can allow students to experience productive struggle without intervening too early, while recognising when additional support is needed. Crucially, this orchestration should include a motivational escalator: a range of timely, personalized interventions that can restore confidence, sustain engagement, and help learners overcome challenges without taking away the opportunity to learn through effort. The goal is not to eliminate struggle, but to keep learners in the zone where challenge drives learning rather than disengagement.
Daniel McCrea
Consider two learners who spend eight minutes on the same question. Both answer incorrectly four times, request two hints and eventually abandon the activity.
The raw behavioural data is identical, but the reasons for abandonment may be very different.
For the first learner, our models indicate strong mastery of prerequisite concepts. The question is unusually difficult, but within their reach. This may be learners experiencing ‘productive struggle’.
For the second, our models identify gaps in prerequisite knowledge and a history of similar errors followed by abandonment. This suggests the learner may be genuinely stuck and needs a different intervention.
We can also explicitly ask learners how they feel when exiting a session, capturing their affective state: were they frustrated, challenged, confident or simply interrupted?
By combining behavioural data, learner modelling and self-reported feedback, we can distinguish between productive struggle and genuine difficulty much more accurately, and intervene accordingly.
Arthur Alexion
Within the context of existing adaptive learning environments, creating productive struggle is strongly linked to assessing a learner's mastery of prerequisite skills. So accurate mapping of prerequisites and accurate assessment of mastery are critical to creating productive struggle.
It’s hard to see the difference between ‘productive struggle’ and ‘genuinely stuck’ directly but serving up recommended learning content based on mastery of prerequisites together with monitoring of outcomes are learning science’s best way to get the distinction right most of the time.
That’s the ‘how’ learning science can get the distinction right. The ‘why’ is set out in our answer to the first question — productive struggle is where learning happens!
3. Most education systems collect enormous amounts of learner data, but very little of it feeds back into how content or assessment is actually adapted. Where have you seen it work in practice?
Jessica White
When learner data is treated as a starting point rather than an answer, it becomes part of the learning process: dynamic and open to change. The risk lies in treating a learner’s data as a fixed diagnosis, profiling learners into positions they cannot move out of. A misdiagnosis that a learner can’t challenge or change can do real harm, which is why the ethics of learner data matters so much.
In one Adaptemy pilot, the early learner data told only half the story. When we triangulated it with classroom observation, think-alouds and conversations with the teacher, it became clear that learners weren’t disengaged. They were experiencing productive struggle and working concepts out in dialogue with their peers, so we reiterated and tried to find ways to integrate these findings into a new approach.
That gap shaped the next round of product and learning design. We developed a Teacher Pedagogical Guide to be embedded in the platform, that brought teachers into the process of designing their curriculum design in response to our learning design. Teachers felt their curriculum design was more integrated with the specificities of how the content was being adapted, based on student’s interactions and responses.
So it was this combination of their adaptive teaching being in tandem and often mirroring adaptive learning from the platform that guided students to the most useful areas and resources when they were stuck, or towards different approaches. That gave both teachers and learners more agency to use the platform effectively for their own context, both individually and with others.
So the data surfaces a starting point which can show a discrepancy or some aspect that seems confusing or incongruent for both learners and teachers, that can lead to further curriculum and content and learning redesign, leading to more targeted learning.
Daniel McCrea
We've seen this work directly with our publishing partners, where learner response data is used both to improve learning content and to make assessments more effective.
For example, our analysis identified videos associated with worse subsequent learner performance. The engine could recommend those resources less frequently while editors investigated the problem.
For assessments, we calculate each question's actual difficulty and how effectively it distinguishes between different levels of learner mastery.
We can then apply the principle of maximum information gain: rather than presenting every learner with the same fixed sequence of assessment, the engine dynamically selects the next question most likely to improve our understanding of what that learner knows. This makes assessments more accurate and efficient, potentially requiring fewer questions to reach a reliable conclusion.
Crucially, these insights are based on sufficient evidence, not isolated responses. Editors receive actionable recommendations and retain control over what changes.
The principle is simple: use learner data not only to personalise learning, but to continuously improve the quality and effectiveness of the content and assessment itself.
Arthur Alexion
Adaptation of content and assessment based on learner data can be built into an adaptive learning platform’s capabilities.
The first thing to say is that it does not require personal profiling of learners. This type of analysis can be done on anonymous learner data and that’s the way Adaptemy’s engine is designed. The data we collect is limited to what we need to extract for insights.
At runtime, an adaptive learning engine is looking to improve how it understands an individual learner but it is also gathering information on what content and assessments work. Does it discriminate well between levels of mastery? Is it confusing? The observations on content can be summarized and fed back to the learning platforms content team as observations and recommendations on content. Some Adaptemy customers use these observations as their primary means of prioritizing platform content improvements.
Ioana Ghergulescu
The key is not how much learner data we collect, but how effectively we turn that data into meaningful learning decisions. In practice, this works best when there is a clear framework connecting data collection, learner modelling, and learning orchestration. Adopting standards such as Experience API (xAPI) enables consistent tracking of learning experiences across content, assessments, and platforms.
Equally important is data minimisation: collecting only the data that is relevant and actionable, rather than accumulating data without a clear purpose. This needs to be supported by robust learner models and well-defined thresholds for the quantity and quality of evidence required before machine learning algorithms update those models. When these elements work together, learner data becomes part of a continuous feedback loop, where insights inform the learner, content and curriculum models and the updated models drive meaningful adaptations to the overall learning experience. The real value comes not from collecting more data, but from closing the loop between evidence and action.
4. If research says that productive struggle improves retention, how does that get translated into something an institution can deploy at scale?
Jessica White
What scales is the learning design that creates the conditions for learners’ productive struggle. This means allowing learners to attempt a question before the explanation and spacing practice throughout their learning journey. Furthermore, scaffolding affective support, such as encouragement to learners that are finding the experience too difficult. Or further challenge to those learners who are ready to tackle more difficult problems, rather than explaining answers too soon. This supports learners’ self-efficacy which can develop life-long learners. It also means building in space for reflection, where learners explain their thinking, so that their productive struggle becomes visible within their own context. This is so that they can understand more specifically where they need to develop.
These patterns and spaces of reflection can be built into a platform once and reach every learner. What doesn’t scale automatically is the classroom/ learning environment around it. That’s why teachers and learners need to be part of the early co-design, and why a teacher’s pedagogical guide can be deployed at scale alongside or embedded within the product. It makes teachers / trainers integral to the solution of designing for learning, rather than the platform being a bolt-on solution that often seems disconnected from their context.
For institutions, this means investing in teachers’ and trainers’ professional development as part of the rollout. It also means making it explicit in the learning design of adaptive platforms that ‘productive struggle’ is part of learning. For example, in ‘productive failure’ sequencing, learners deliberately attempt a problem before they are taught how to solve it. Explaining this step-change is part of scaling. Otherwise, learners and teachers may mistake the challenge for a design fault.
Daniel McCrea
The research is well established, but a simple exercise makes it real. Think about the last ten YouTube videos you watched. How many can you remember? More importantly, how much of what you watched could you actually apply tomorrow?
Watching something is not the same as learning it. Learning requires effort: recalling information, applying knowledge, making mistakes and working through challenges. That productive struggle is what helps learning stick.
The challenge is creating those moments of productive struggle for every learner, at scale. This is where sophisticated learner modelling becomes essential. An adaptive learning engine builds a probabilistic model of each learner's mastery, estimating what they already understand and how likely they are to succeed with a particular piece of content or assessment at a given level of difficulty.
If those predictions are sufficiently accurate, we can consistently select activities that sit right at the edge of an individual learner's capability: challenging enough to require genuine effort, but achievable enough to avoid frustration. As learners respond, the model updates its estimates and adjusts the level of challenge accordingly.
That's how we translate learning science into technology: by making productive struggle something we can deliberately create for millions of individual learners, rather than leaving it to chance.
Arthur Alexion
Productive struggle at scale requires automation. It can only be achieved by a digital (adaptive) system.
Adaptive learning creates productive struggle by understanding individual learners and understanding individual pieces of content and assessment and recommending the right ‘next step’ in the learning journey. As mentioned previously, having well defined prerequisites is critical but also having a well thought through syllabus is also critical. With the right ‘guidance’ from prerequisites and syllabus, the adaptive learning technology will allow effortless, massive scaling.
Ioana Ghergulescu
The key is to translate learning science into the learning design and orchestration of a truly adaptive product, rather than treating productive struggle as a standalone pedagogical principle. This means designing learning experiences that deliberately create the right level of challenge, supported by accurate learner models, content difficulty estimates, and intelligent interventions that recognise when to let learners persist and when to provide support.
Crucially, the product must close the loop between learner behaviour, evidence, and adaptation, continuously refining its understanding of each learner and adjusting the learning experience accordingly. When these principles are embedded into the product architecture and learning design, institutions can deliver the benefits of productive struggle consistently and at scale, rather than relying solely on individual educators to implement them.
5. University leaders are being asked to invest in AI-powered tools but also to explain to faculty and students how those tools work. What does it take to make an adaptive learning system transparent enough that educators and students actually trust it?
Jessica White
Trust comes from people being able to see and question what an adaptive learning platform processes. This can be by way of analytics. An open learner model, showing learners and teachers what the system represents the learners as knowing, turns a hidden judgement into something they can reflect upon, interpret in their own way and be used as a mediation tool, rather than a rule.
What makes difference in practice: having explainability in the platform as to why a task was recommended, letting teachers/ trainers override system’s decisions in their own ways of using the platform and that relates to contextual flexibility.
Finally, being honest that the platform cannot solve every learning or teaching challenge. However, integrated well, inclusively and ethically within teaching and learning practices, as well as allowing space for learners and teachers/trainers to have room to support further iteration and co-design, can go a long way to supporting many challenges we face.
Research on self-regulated learning and metacognition shows learners engage more deeply and holistically, when they understand the context of their learning, how it fits into the bigger picture and how it is relevant to their lives. Transparency and some explainability, isn’t just ethical; it’s good pedagogy.
Daniel McCrea
“You cannot trust a system that you do not understand.” When we were a younger company, we thought customers wanted to hear “don’t worry about the magic box, just look at the results.” We were wrong: explainability creates trust and trust is critical. Trust requires three things: understanding how the system makes decisions, being able to explain individual recommendations, and independently verifying that the technology works.
For institutional leaders, that means transparency about the underlying models, their configuration and their effectiveness. We share underlying data with our partners so they, or independent researchers, can evaluate our models themselves. For teachers and students, transparency means understanding what the system believes a learner knows, what it recommends next, and why.
No educational institution should have to trust a black box. The models should be explainable, the recommendations understandable, and the outcomes independently measurable.
Arthur Alexion
Trust can come from two sources: ‘do your own research’ and ‘trust the experts’.
In terms of allowing faculty and students to gain direct confidence in adaptive learning, this is possible only with non-black box solutions. Using Claude or ChatGPT to directly drive the learning path is not compatible with direct confidence. This type of approach is not deterministic — the same scenario can lead to a different recommendation because the underlying intelligence is probabilistic. Adaptemy’s technology deliberately avoids this approach by using Bayesian probability and other mathematical methods which are transparent and therefore explainable. Obviously, the maths is complex so anyone wanting direct understanding needs to have a certain level of mathematical knowledge.
Alternatively, trust can come from expert assessment of the effectiveness. There is a body of research evidence of generic research which concludes positively on the impact of adaptive (AI-powered) learning; and if needed, universities can commission.
Ioana Ghergulescu
Trust in adaptive learning systems requires transparency not only in the decisions AI makes, but also in the models and evidence behind those decisions. This starts with the learner model: making visible what the system estimates a student knows, what evidence supports those estimates, and how confident the system is in them. The same principle applies to other models, including content and curriculum models, where estimates and their supporting evidence should also be transparent.
Equally important is explainability in learning orchestration, helping learners and educators understand why particular content, assessments, or interventions are recommended. But transparency alone is not enough. We also need to enable meaningful human-AI collaboration, where learners can provide context about their immediate goals, such as whether they are learning something new or revising, and the system takes that into account. Similarly, teachers and lecturers should be able to contribute their knowledge of the learner and the learning context to inform and guide AI-driven decisions. Ultimately, trust comes from making AI not only explainable, but also collaborative, allowing learners and educators to understand, influence, and shape the learning experience.
An Insightful Conversation
The Adaptemy team agrees on the destination: AI that responds to each learner's needs. But they arrive at it from different perspectives. Ioana Ghergulescu starts from the learner model and works outward to the product. Jessica Kennedy White starts from how learners think about their own thinking. Daniel McCrea talks about what happens after the pilot, when the real work of adoption begins. Arthur Alexion offers his views on the mathematics that makes the system explainable.
The team's recurring concept of "productive struggle" is a useful test for any adaptive system. Does the technology give learners space to think, or does it optimize for ease? And can it tell the difference between the two?
The technology should start from how people learn and work toward the product, not the other way around.
Coming in Part 2
In Part 2, our guests into the operational questions: why completion rates and satisfaction scores are misleading most provosts, which data safeguards institutions overlook until it's too late, what eleven years of working with younger learners taught them about screen time, and what each of them would change about how education leaders are approaching AI right now.
EDT&Partners
The EDT&Partners Editorial Team brings together education and technology experts sharing insights, stories, and strategies shaping the future of learning.
Join our newsletter
Be part of our global community — receive the latest articles, perspectives, and resources from The EDiT Journal.
