Baillifard, Gabella, Lavenex & Martarelli, (2025) reported on a study on integration of AI tutors to complement a learning program and reported that students who used the AI tutor achieved significantly higher grades. The researchers also evaluated the effectiveness of the AI-tutor’s neural-network models to model human learning processes. The name of the company who built the app was provided and, at a high level, how it is used but not details on how the app was built nor details on how it works. Students are presented with questions and based on interactions with students and their answers to questions, the AI tutor dynamically predicted the probability of a correct answer – referred to as the ‘grasp’ – for each student and each question. The questions encompassed various formats such as definitions, clozes (fill-in-the-blank), true/false, multiple-choice, image-based, and acronyms. The questions also had varying difficulty levels to accommodate the proficiency distribution among students. Each question was linked to specific course materials to provide students with contextual feedback if needed. I agree, we could and should make use of Universal Design for Learning (UDL) and differentiation in the design of an intelligent tutoring system (ITS).

This reminds me of MathXL from Pearson and ALEKS from McGraw Hill.

The ‘grasp’ for each student and each question is the probability of a correct answer and was dynamically predicted based on interactions with the students and their answers to questions.

These predictions were made by an artificial neural network trained with input features derived from information on questions, students, and their historical interactions. With this personalized understanding of each individual student’s knowledge levels and their evolution (through learning and forgetting),the AI tutor presented the questions considered most relevant and beneficial whenever the student accessed the app. The questions that were selected by the app aimed to maintain an appropriate level of challenge for each student, avoiding unstimulating easy as well as frustratingly hard questions. This approach aligns with the concept of “desirable difficulty” and ensures an engaging learning experience within the student’s zone of proximal development according to Lev Vygotsky formulation of how humans learn.

The app provided students with the capability to track their progress through a visual representation of their knowledge called the “learnet”. The learnet presented a three-dimensional organization of all the key concepts from the learning materials and their interrelations. Each point of the learnet corresponded to a specific course concept, with its brightness indicating the student’s grasp of that particular concept. Darker learnets served as motivation for students, indicating areas where they still had more to learn or had already forgotten. Conversely, brighter learnets communicated to students their high knowledge levels.

The authors provided high-level background information on human learning, machine learning, and using machine learning to enhance human learning. Here I gather some of the useful information. Some of the most effective techniques for human learning: spaced practice, retrieval practice, interleaving, elaboration, and personalization.

Personalization aims to create flexible learning experiences tailored to meet the unique needs of each student and the approach that interests me the most. In my readings and my studies, I read that implementing personalization is not easy, which is what Baillifard et al. (2025) reported here also. Personalization advantageously combines these techniques by determining the appropriate spacing levels for each learner, by selecting the levels of difficulty for retrieval practice exercises, and by interleaving and connecting concepts in a way that is adapted to individual progress. The desired difficulty depends on individual capacities, preferences, and energy levels, which aligns neatly with teaching students not motivated to learn or neurodivergent students with, for example, a learning disability, which is what interests me.

Baillifard et al. (2025) wrote about applications of AI in education grouped into four main categories: (1) profiling and prediction (2) assessment and evaluation (3) adaptive systems and personalization (4) intelligent tutoring systems. The authors provide an overall observation that AI technologies serve as valuable assistants, providing students with personalized support in their zone of proximal development, where they can rapidly develop with appropriate assistance at any time of the day or week.

This is instructive for me because prior to now I thought of adaptive systems and personalization and intelligent tutoring systems as being in one category. For me, my interests remain finding out more about intelligent tutoring systems teaching students not motivated to learn or neurodivergent students with, for example, a learning disability.

Kulik & Fletcher (2016) confirmed for me to go with the name intelligent tutoring systems (ITS). The authors found that ITSs can be effective instructional tools and can have a significant effect on improving student’s performance. I found it interesting that the teacher’s experience with ITS instruction is a factor on the effectiveness of ITS on students performance. The authors reported that a teacher who was experienced with ITS instruction resulted in improved performance by the students whereas teachers with limited experience with ITS resulted in a negative effect on student’s performance. This is significant for me because it raises a question for me: would students be able to achieve significant gains if they were to work independently with an intelligent tutoring system?

Beverly Park Wolf (2009) started by explaining what an inflection point is and argued that technology has generated an inflection point in education with three drivers of this educational inflection point being AI, cognitive science and the Internet. Wolf (2009) claimed that cognitive science and AI are two sides of the same coin which means that ultimately leaves AI techniques and the Internet. Wolf (2009) reported on intelligent tutors which is the same as intelligent tutoring systems (ITS). The author explained that intelligent tutoring systems (ITS) are student-centered; students move at their own pace, obtain their own knowledge, and engage in self-directed learning or group-directed learning. Intelligent tutoring systems contain rich, dynamic models of student knowledge that depict the key ideas learners should understand as well as common learner conceptions and misconceptions. They have embedded models of how students learn and teachers teach and can adapt their model over time as student’s understanding becomes more sophisticated.

The quote “What one knows is, in youth, of little moment; they know enough who know how to learn.” Henry Adams (1907) caught my attention and held my interest because it is so appropriate for what the author is saying about education.

I realized that I need to take notes before I can make notes. I came to this realization while I was reading Wolf (2009). I discovered that I needed to take notes using pencil and paper (and not a pen) while I was doing the reading in order for me to make notes in my journal.

Wolf (2009) made the argument that there is an inflection point in education driven by AI, cognitive science and the Internet. I do not disagree with what cognitive science does, nor do I disagree that the Internet provides an unlimited source of information, available anytime anywhere. I do not know what to believe about Wolf (2009) statement regarding AI as one of the drivers, if we mean ChatGPT and generative AI.

I do not argue with Wolf (2009) statement that AI, the science of building computers to do things that would be considered intelligent if done by people, leads to deeper understanding of knowledge, especially representing and reasoning about “how to” knowledge, such as procedural knowledge.

Michael Wagner (2024) presents how AI is changing education but he means generative AI, However, as Steven Pemberton pointed out, generative AI are just stochastic parrots (SAI Conference, 2024).

On the other hand, as I read Wolf (2009) a second time, I believe that Wolf (2009) is not only talking about generative AI. Wolf (2009) states that AI and cognitive science are both about understanding the nature of intelligence; AI techniques are used to build software models of cognitive processes and results of cognitive science are used to build more AI techniques to emulate human behaviour. Wolf (2009) adds that AI techniques are used in education to model student knowledge, academic topics, and teaching strategies.

Wolf (2009) defines intelligent educational software as those that use intelligent techniques to model and reason about learners, including simulations; advisory, reminder, or collaborative systems; or games. Wolf (2009) specifically mentioned intelligent tutors and stated that these are based on student strategies. Intelligent tutors, according to Wolf (2009), contains rich, dynamic models of student knowledge that depict the key ideas learners should understand as well as common learner conceptions and misconceptions. They have embedded student models that reason about how people learn, specifically how new knowledge is filtered and integrated into a person’s existing cognitive structure.

Then Wolf (2009) states that intelligent tutors have embedded models of how students and teachers reason and can adapt their model over time as students’ understanding becomes increasing sophisticated. With intelligent tutors, students move at their own pace, obtain their own knowledge, and engage in self-directed learning or group directed learning.

This gave me paused, because I thought intelligent tutors are about students and yet Wolf (2009) include how teachers reason.

In response to the question in this module “What are Intelligent Tutoring Systems (ITS)?” and “What are Adaptive Learning Systems?”, felt that I had an “Eureka!” moment in that my understanding in this area was clarified in this module. This is because I have been interested in personalized learning and adaptive learning systems for a number of years now prior to this course, going all the way back to 2020. For example, I am familiar with ALEKS https://www.aleks.com/about_aleks before this course.

I remembered “Ding!” that we talked about in this course as I read Wolf (2009) describe intelligent tutors as having fantastic abilities and capabilities. Then Wolf (2009) describe AI tutors as having equally fantastic abilities and capabilities. I feel the need to copy these descriptions in my journal for future reference.

“AI tutors work with differently enabled students, make collaboration possible and transparent, and integrate agents that are aware of students’ cognitive, affective, and social characteristics. Intelligent agents sense, communicate, measure, and respond appropriately to each student. They might detect learning disability and modify the pace and content of existing pedagogical resources. Agents coach students and scaffold collaboration and learning. They reason about student discussions, argumentations, and dialogue and support students in resolving differences and agreeing on a conclusion. They monitor and coach students based on representations of both content and social issues and reason about the probability of student actions. Probability theory (reinforcement learning, Bayesian networks) defines the likelihood of an event occurring during learning. AI techniques contribute to self-improving tutors, in which tutors evaluate their own teaching” (Wolf, 2009, p.8).

Wolf (2009) describes the state of the art in AI and education (AIED) in 2009. One of the goals was to develop software that captures the reasoning of teachers and the learning of students. This process begins by representing expert knowledge (e.g. as a collection of heuristic rules) capable of answering questions and solving problems presented to the students. For example, an expert system inside a good algebra tutor represents each algebra problem and approximates how the “ideal” student solves those problems. Student models, the student systems inside the tutor, examines a student’s reasoning, find the exact step at which the student went astray, diagnose the reasons for the error, and suggest ways to move the student forward.

Wolf (2009) discussed the potential value of automated intelligent tutors. First, that one-to-one tutoring is the best way to learn. So the Holy Grail of teaching technology is each student having their own automated tutor capable of finely tailoring student learning experiences to students’ needs.

We see Sal Khan saying the same thing in 2024, see Module 3 of this course and I mentioned it in the next post in my Learning Portfolio here.

Wolf (2009) identified some of the goals of AI and education (AIED). Among these are providing students with alternative representations of content, alternative paths through material, and alternate means of interaction. Another goal is to understand how human emotion influences individual learning differences and the extent to which emotions, cognitive ability and gender impact learning. Some larger research questions are:

What is the nature of knowledge and how is it represented? How can an individual student be helped to learn? Which styles of teaching are effective, and when should they be used? What misconceptions do learners have?

Wolf (2009) identified some visions of AIED in 2009. One such vision is “a teacher for every student” or “a community of teachers for every student. This vision includes making learning a social activity accepting multimodal input from students (handwriting, speech, facial expressions, body language) and supporting multiple teaching strategies (collaboration, inquiry, and dialogue).

Wolf (2009) imagined futuristic scenarios when such perfect automated intelligent tutors are available anytime from any place, and on any topic.

I have a need to keep these paragraphs for my future reference so I believe that it is beneficial and efficient that I copy these paragraphs from Wolf (2009) into my Learning Portfolio here.

“Intelligent tutors know individual student differences. Tutors have knowledge of each student’s background, learning style, and current needs and choose multimedia material at the proper teaching level and style. For example, some students solve fraction problems while learning about endangered species; premed students practice fundamental procedures for cardiac arrest; and legal students argue points against a tutor that role-plays as a prosecutor” (Wolf, 2009, p.9).

“Such systems infer student emotion and leverage knowledge to increase performance. They might determine each student’s affective state and then respond appropriately to student emotion. Systems recognize a frustrated student (based on facial images, posture detectors, and conductance sensors) and respond in a supportive way with an animated agent that uses appropriate head and body gestures to express caring behavior” (Wolf, 2009, p.9).

“Such systems can also recognize bored students (based on slow response and lack of engagement) and suggest more challenging problems. Intelligent tutors work with students who have various abilities. If a student has dyslexia, the tutor might note that he is disorganized, unable to plan, poorly motivated, and not confident. For students who react well to spoken text messages, natural language techniques simplify the tutor’s responses until the student exhibits confidence and sufficient background knowledge. During each interaction, the tutor updates its model of presumed student knowledge and current misconceptions” (Wolf, 2009, p.9).

“Students work independently or in teams. Groups of learners, separated in space and time, collaborate on open-ended problems, generate writing or musical compositions, and are generally in control of their own learning. In team activities, they work with remote partners, explaining their reasoning and offering suggestions. They continue learning as long as they are engaged in productive activities. Teachers easily modify topics, reproduce tutors, at an infinitesimal cost to students and schools and have detailed records of student performance” (Wolf, 2009, p.9).

“Intelligent tutors know how to teach. Academic material stored in intelligent systems is not just data about a topic (i.e., questions and answers about facts and procedures). Rather, such software contains qualitative models of each domain to be taught, including objects and processes that characterize trends and causal relations among topics. Each model also reasons about knowledge in the domain, follows a student’s reasoning about that knowledge, engages in discussions, and answers questions on various topics. New tutors are easily built and added onto existing tutors, thus augmenting a system’s teaching ability. Tutors store teaching methods and processes (e.g., strategies for presenting topics, feedback, and assessment). This knowledge contains rules about how outstanding teachers behave and teaching strategies suggested by learning theories” (Wolf, 2009, p.9).

Wolf (2009) talked about effective teaching methods. First, that one-on-one tutoring by human teachers is the most effective teaching method. Other effective methods are collaboration, inquiry and teaching metacognition.

Wolf (2009) concluded this chapter with a number of assertions, a few that interests me the most. The ability of automated intelligent tutors to enhance though not replace one-to-one human tutoring, and the potential for extending teaching and learning methods, e.g. collaborative and inquiry learning.

If automated personalized tutoring can be made widely and inexpensively available, intelligent tutors have the potential to provide a skilled teacher, or a community of teachers for every student, anywhere, at any moment.

Chapter 2 of Wolf (2009) deals with features of intelligent tutors. This chapter begins with the statement that building intelligent tutors requires understanding teaching and learning. Wolf (2009) then discusses the features of intelligent tutors with the caveat that few features (at that time of writing) contain all these features and more research is needed to achieve these features fully.

The first feature of intelligent tutors in generativity; the ability to generate appropriate resources for each student. Generativity is the ability to generate customized problems, hints, or help based on representing subject matter, student knowledge, and human tutor capabilities.

The second and third features of intelligent tutors are student knowledge (dynamically recording learned tasks based on student actions) and expert knowledge (representing topics, concepts, and processes of the domain).

The fourth feature of intelligent tutors is mixed-initiative, or the ability for either student or tutor to take control of an interaction. True mixed-initiative enables students to ask novel questions and set the agenda and typically requires generating natural language answers.

At the time of writing, Wolf (2009) reported that most existing intelligent tutors are mentor-driven, e.g. they set the agenda, ask questions and determine the path students will take through the domain.

The fifth feature of intelligent tutors is interactive learning or being responsible for students’ learning needs.

The sixth feature of intelligent tutors in instructional modelling, or how a tutor modifies its guidance for each student. Instructional modelling means receiving input from the student model, because students with less prior knowledge clearly require more instruction and guidance than students with more knowledge.

The seventh feature of intelligent tutors is self-improving or modifying a tutor’s performance based on its experiences with prior students. This feature is often implemented through machine learning and data mining techniques that evaluate prior students’ learning experiences, judge which interventions are effective, and use this information to change tutor responses.

Just a few quick points from the video podcast. Dan Schwartz said he believes that AI will not fully replace a human being but it will be more about augmentation. People will use AI to do more, to use AI to do things they could not do before. Denise Pope said that AI will not replace radiologists, rather radiologists who use AI will replace radiologists who don’t use AI. Victor Lee reminds us that AI is not human. It can do things that appear human-like but it is still not human. Humans make life important critical judgements and we need to make sure that humans are still doing that.

References

Baillifard, A., Gabella, M., Lavenex, P. B., & Martarelli, C. S. (2025). Effective learning with a personal AI tutor: A case study. Education and Information Technologies, 30(1), 297–312. https://doi.org/10.1007/s10639-024-12888-5

Kulik, J. A., & Fletcher, J. D. (2016). Effectiveness of Intelligent Tutoring Systems: A Meta-Analytic Review. Review of Educational Research, 86(1), 42-78. https://doi.org/10.3102/0034654315581420

SAI Conference (2024, Dec 11). There’s no I in AI. Join Steven Pemberton as he delves into the fascinating world of Artificial Intelligence and uncovers the latest advancements and trends that are shaping the future of humanity. [Video]. YouTube. https://www.youtube.com/watch?v=lS4-QSR1sNk

Schwartz, D., & Pope, D. (Hosts). (2024, Oct 22). On this episode of School’s In, hosts Dan Schwartz and Denise Pope welcome Associate Professor Victor Lee as they discuss the rise of artificial intelligence (AI) in education and its implications for how we teach and learn. [Video podcast episode]. Stanford Graduate School of Education. https://www.youtube.com/watch?v=-2nt-f87478

Wagner, M.G. (2024, Sept 27). How AI Is Changing Education: Time to Take Action. [Video]. YouTube. https://www.youtube.com/watch?v=3sHiNTJ9Qsg

Wolf, Beverly Park. (2009). Building intelligent interactive tutors : Student-centered strategies for revolutionizing e-learning. Elsevier.