Pragmatic AI in education and its role in mathematics learning and teaching

Figure 1 also indicates the potential for AI (the robot icon) to contribute to the development of the underlying educational goals (SEL development in this example), as well as assisting in mediating both the balancing and reinforcing feedback loops (B1 and R2) associated with emotions. Some of these envisaged points for interaction are autonomous, where the robot icon appears alone, and some are supervised, where the robot and mortarboard coexist. It is important to note that these potential points of AI support are entirely indicative and the examples that are presented here are general suggestions.

Before exploring Fig. 1 in detail, it is important that we emphasise the cyclical nature of this model. Learning does not happen as a one-off event. Rather the objects of education (SEL development in this example) result in a series of changes in other quantities that are then available for use in future cycles. It is also essential to recognise that the system presented in Fig. 1 is incomplete and forms just one sub-system in a much more complex process. That is, there will be other factors in the larger system connected to the nodes of our model that are not represented here (e.g., resilience). For this reason, it is important that students experience both successes in learning and difficulties. Each will cause different emotional responses, and the larger system will respond in different ways potentially developing attributes like perseverance that are not represented here. Any AI implementations that become incorporated into this system need to ensure that they are designed to respond appropriately to both success and difficulty in ways that enhance the underlying educational objectives without disempowering students or their teachers or shortcutting the learning processes.

Working from our conceptual model (Fig. 1), we propose six general areas where AI may be able to provide mechanisms or approaches that support the learning process. From these we have identified specific elements of learning design that may be targeted for research, improvement, or modification. These objects of transformation and their associated design goals are presented in Table 1 along with possible AI application strategies and examples of implementation.

Firstly, supporting teachers to provide mathematical learning activities that are personalised at an appropriate level of complexity to match the cognitive abilities of individual students is essential. The strengths of current approaches to AI lie in their ability to mimic human processes, using vastly larger quantities of data, and with far greater speed that humans can achieve. When an AI is presented with a problem that is similar to ones it has ‘seen’ before, then it is reasonable to expect the algorithm to follow a similar process in generating output. Therefore, the implementation of some form of semi-supervised, or even unsupervised, algorithm to analyse each student’s past and current learning performance and/or mastery data and to suggest learning pathways for them to follow or to dynamically adjust the task demands to an appropriate level would be a potentially effective use for an AI learning support. Such an implementation would appear at ① in Fig. 1 and provide personalised challenges that maintain an optimal balance between task demand and cognitive quality of instruction. These individualised learning experiences will align with each student’s proficiency level, minimising their frustration and enhancing the likelihood of successful learning outcomes. However, personalisation should not be limited to task difficulty alone. AI can help pinpoint specific moments where students struggle, providing feedback that is not only task-focused but also supports positive appraisals of their efforts. For example, AI-driven systems can identify when a student shows signs of frustration or disengagement and offer supportive feedback that highlights their progress and effort rather than solely focusing on task completion. This can help students reinterpret their experiences as opportunities for learning, thereby fostering a more positive self-concept and increasing their motivation. An algorithm such as this would be best suited for general implementation by resource producers, as they would have ready access to the large amounts of data needed to establish the parameters of the algorithm. However, in a semi-supervised arrangement, some system or process would need to be developed to allow the classroom teacher to provide input and fine tune the algorithm in a straightforward and intuitive way.

Extending the idea of performance monitoring to student self-monitoring, AI algorithms can be used to enhance learners’ perceptions of agency and control and might be utilised at ② in Fig. 1. Using a large set of learning performance data a resource producer could develop a predictive model that gives students a selection of recommended next steps for their own personalised learning yet remains within the same area of knowledge that the teacher has assigned. In this conception, the AI algorithm essentially creates a decision tree for building learning pathways but leaves the final step of the process, the decision itself, up to the student. Built into such a predictive system would be the opportunity for learning analytics that could build an understanding of each student’s cognitive strengths and weaknesses and provide nudges to assist the student in achieving their own stated goals. AI-enabled tools might also be used in creative learning environments encourage cooperation and thus cultivating positive social interactions. These tools might include interactive and collaborative platforms where students can explore mathematical concepts independently and engage in cooperative problem-solving activities. AI can also contribute to the development of enhanced feelings of control and growth through individualised achievement goal structures and expectations; approaches that are key in reducing maths anxiety.

AI may also assist teachers in enhancing students’ value induction ③ which is an important facet of learning success. AI search engines have access to an almost endless volume of information that could be used to select contexts or situations that highlight the real-world applications and uses of mathematical problems in a way that is relevant to individual students. Such an approach can contribute to students’ understanding of the intrinsic value of mathematical knowledge¹⁵.

Chatbots based on large language models (LLMs) have become particularly effective in recent years. While chatbots are still far from perfect, the ability to sideload an LLM with an appropriate set of background data files, does offer educators the possibility to use these algorithms in a safe and closed environment while ensuring that the chatbot has access to only appropriate additional information and not the entire unfiltered web. Adopting such an approach has the added advantage that the LLM itself can be smaller allowing it to run on prosumer level local hardware and removing the need to run on massive cloud infrastructure. The use of an AI approach like this to assist students in moderating feedback loops, such as the one at ④, has great potential for research and impact¹⁶. As noted, there is a reinforcing loop between perceptions of control and maths anxiety (R1 in Fig. 1). In a traditional classroom, this feedback loop might be driven by general reflection on the task and its success criteria analysing what the student got wrong and how to fix this. This type of feedback needs to be carefully supervised to ensure that unhelpful self-talk does not dominate the process. An AI chatbot can help to reframe these reflections to consider how an individual has developed in multiple ways while engaging with the learning activities and reduce the dichotomous focus on success vs. failure.

Furthermore, retraining students’ notions of failure and success becomes feasible through AI interventions that emphasise the iterative nature of learning and the value of the underlying educational goal. Such AI tools placed at ⑤ could promote the use of personal micro-targets for students and shift the focus of the cognitive process away from the learning activity and onto this deeper learning goal, AI-driven interventions can lead to a shift in student mindset as they learn to view setbacks as opportunities for growth and reduce the anxiety associated with performance.

However, while current AI tools are capable of ‘decision making’ using large amounts of mostly structured data, they lack the capability to genuinely understand and adapt to the emotional and cognitive needs of students. The AI models tend to focus more on the mechanics of teaching rather than the emotional well-being of the learner. That is, currently available AI tools are more focussed on the products of learning than on the human process of learning. To truly harness the potential of AI in transforming mathematics education, we need to prioritise different research goals. Researchers should therefore aim to develop new systems and approaches that foster a deeper, more intuitive understanding of mathematical concepts, rather than just improving the efficiency of content delivery.

For example, we know that self-regulation of emotions is essential in order to manage adverse situations such as maths anxiety. Natural language processing algorithms might be developed to identify the fingerprints of both positive and negative emotional meaning in students’ extended textual responses. This might be supplemented with data gathered through multi-modal learning analytics—such as computer vision and audio analysis—to track non-lingual cues such as facial expressions, tonal variation and gestures¹⁷. Together such a dataset might be able to be used to provide support and strategies, both virtual and real-life, that can assist students in developing their self-regulation and emotional moderation. Such a technology might have a place at ③ or ⑥ in Fig. 1.