Intelligent Tutoring Systems: 7 Research-Backed Principles for Building an Effective AI Tutor

After spotting his LinkedIn post on intelligent tutoring systems, we asked longtime Third Space Learning contributor Neil Almond, founder of the AI newsletter Teacher Prompts, to share how he believes ITS, such as AI tutors, should be grounded in the best research-informed practice.

Below, he outlines the essentials that make any ITS genuinely effective and ideally as effective as traditional intelligent tutoring.

For each of the principles below, we also show what this looks like in practice using the intelligent tutoring system example of Skye the AI maths tutor created by Third Space Learning and now used by hundreds of pupils in schools across the country.

What are intelligent tutoring systems?

Intelligent tutoring systems (ITS) are AI-driven programs that diagnose each learner’s knowledge state in real time. They then deliver scaffolded hints, questions and feedback until mastery is reached. The best artificial intelligence tutoring systems use intelligent computer assisted instruction to replicate the benefits of one-to-one human tutoring.

Delivering on the promise of intelligent tutoring systems

The education sector is feeling the rapid evolution of artificial intelligence. Generative AI has accelerated the development of intelligent tutoring systems that promise personalised education at scale.

When implemented thoughtfully, these systems do, of course, hold promise for enhancing the learning process within many subjects. From maths to improving writing skills, both in and out of the traditional classroom setting. 

With the financial squeeze on schools, intelligent tutoring systems provide a real opportunity to make the benefits of one to one tutoring available for all students who need it. 

But, for intelligent tutoring systems to truly deliver on their potential of genuine personal tutoring, it must rest on strong research and account for the human factors. It must be built on a solid foundation of well-researched pedagogical principles.

Having reviewed various generative artificial intelligence learning tools that claim to support pupils’ learning, there are some basic elements that I think intelligent tutoring systems, AI tutoring, learning platforms, and similar AI tools should consider as standard for their product development.

These are what I consider the seven principles of effective intelligent tutoring systems.

7 principles of effective intelligent tutoring systems

1. Expertly crafted curriculum
2. Scaffolded learning
3. Active recall and spaced repetition
4. Interleaving for conceptual understanding
5. Encouraging problem generation
6. Error correction and reflection
7. Cognitive load management

1. Comprehensive curriculum crafted by experts

A robust curriculum is fundamental to any effective tutoring system’s approach. Experienced teachers and subject experts should design the curriculum for generative AI to be truly valuable. Human teachers’ expertise should guide the student model, domain model, pedagogical module and user interface that together form the classic four interacting components (sometimes called the interacting components) of an ITS.

While some argue that artificial intelligence should form the custom learning pathway itself, my experience shows that only human teacher-designed curricula guarantee that the content is accurate, coherent, and relevant to the learning context.

Right now, many of the tasks that generative AI excels at involve creating final products that experts would produce. Hence, the proliferation of AI maths chatbots and “AI maths solvers” answering students’ maths homework questions in one go. That is not supporting the learner’s advancement like effective AI maths tutoring does.

It’s not just about the topic progression

A maths curriculum developed by seasoned educators who truly understand the subject’s nuances can incorporate topic progression and guide the student towards relevant applications and pedagogical choices at each stage.

To take one example, when teaching an abstract concept like multiplying fractions, an AI-generated curriculum would be unlikely to incorporate the use of Cuisenaire rods.

However, an educator-led curriculum might well start a multiplying fractions lesson with Cuisenaire rods to make the concept more understandable and connected to the real world for learners. This depth and contextual grounding help students link theory with practice, leading to a deeper understanding and better knowledge retention.

What the research says about curriculum design

Research in curriculum design highlights the importance of expert involvement.

• Biggs and Tang’s (2011) work on constructive alignment underscores the need to align learning outcomes with teaching methods and assessment strategies.
• Darling-Hammond (2010) suggests that curricula developed by educators with expertise in both subject matter and teaching methods tend to result in improved student outcomes.

This research reinforces that an AI tutor grounded in a strong, expert-designed curriculum is significantly more likely to support effective learning.

ITS example – curriculum design in practice

• Experienced human teachers create all maths lessons, not Skye.
• The lesson slides and teaching sequences that Skye teaches have been adapted from the same lessons used with 160,000 pupils receiving traditional tutoring.
• Human tutors have taught many of these maths lessons thousands of times. Each lesson is reviewed and revised for pupils to make the best possible progress.
• To ensure relevance for each child’s context, each online maths lesson aligns with a specific goal (e.g. bridging Year 5 gaps or practising SATs-style questions).
• Skye’s combination of system prompts and per-lesson prompts is based on insights from over 2,000,000 hours of one-to-one online tutoring Third Space Learning has already delivered.

2. Scaffolded learning with gradual hints

Effective learning is about pupils thinking hard about what they are learning. However, a default behaviour of AI-powered tools is to simply provide the answer. Intelligent tutoring systems that give students instant answers when they ask for it denies pupils the cognitive process involved with thinking hard about their learning.

Instead, like experts, effective intelligent tutoring systems need to meet the pupil where they are and scaffold questions or resources carefully to help students learn. Generative AI systems should aim to mirror this teaching strategy by offering gradual hints rather than just providing direct answers.

For instance, if a student is struggling with a challenging multi–step word problem, the machine learning model could initially give subtle cues about potential errors in their approach or logical inconsistencies.

Scaffolding and personalisation support progress to mastery

By incorporating scaffolded learning, intelligent tutoring systems can effectively support the learning pace of students as they move towards mastery.

As with human tutoring, it’s the personalisation of a lesson to the child’s gaps, plus the support to extend their learning, that can engender real progress – especially in maths.

This is where one–to–one online maths tutoring really has its benefits, as many classroom teachers will attest to how difficult it can be to adapt and support the gaps in a class of 30.

What the research says about scaffolded learning

Scaffolding is well-established in educational research.

• Vygotsky’s (1978) theory of the Zone of Proximal Development (ZPD) shows that learners can achieve higher levels of understanding when given appropriate guidance.
• Wood, Bruner, and Ross (1976) explain how scaffolding helps students bridge the gap between what they can do on their own and what they can achieve with assistance.
• Rosenshine (2010) lists scaffolding as one of his 10 principles of instruction on the basis that it reduces cognitive overload by limiting the amount of new information given to students at one time.  

ITS example – scaffolding in practice

• If a pupil answers incorrectly, Skye gives targeted hints (rather than simply marking the question wrong). 
• There are three layers of hints, with each one pinpointing the misconception and nudging the pupil towards deeper understanding.
• Lessons use a consistent “I do, we do, you do” teaching strategy, gradually moving pupils from guided examples to independent problem–solving.
• Because Skye is voice–based, it listens to pupils’ spoken answers and can immediately identify the specific misconception and respond accordingly.
• Adaptive teaching is built into Skye at a system level; Skye uses formative assessments during the lessons and can increase the level of help where needed (or reduce scaffolding when the pupil is on track).

3. Active recall and spaced repetition

Active recall means actively retrieving information from memory, while spaced repetition involves scheduling reviews of that information at increasing intervals. Together, these teaching strategies help consolidate learning and reduce the burden on working memory as it encodes new material.

An effective intelligent tutoring system must draw on machine learning and machine learning program routines that analyse student performance and schedule practice within an adaptive learning environment.

Intelligent tutoring systems can track what each pupil has learned and mastered and quiz pupils on past material at the appropriate stages – regularly at first, then at greater spaced intervals. This approach helps with memory consolidation and helps move information from working memory to long-term memory.

Personalised retrieval plan

For example, when teaching number bonds to 20, the intelligent tutoring system can create a personalised retrieval and revisiting schedule for each pupil based on data such as accuracy and fluency in solving questions that directly or indirectly use these bonds. Keeping track of spacing schedules is difficult for teachers, so this can be a great strength of AI-based tutoring systems for teachers and students alike.

What the research says about active recall and spaced repetition

A strong body of research supports the importance of active recall and spaced repetition for long–term retention.

• Karpicke and Roediger (2008) provide compelling evidence for the effectiveness of retrieval practice, showing that students who actively recall information perform significantly better in tasks requiring long–term retention.
• Cepeda et al. (2006) demonstrate that spaced repetition can significantly improve memory retention over time.

Incorporating these evidence–based strategies into AI tutoring systems ensures that learners can retain knowledge more effectively, building a solid foundation for future learning.

ITS example – recall and spacing in practice

• Skye’s curriculum is sequential. Lessons break down into manageable steps that frequently revisit previously taught skills. This ensures ongoing retrieval practice.
• The structure of lessons within each programme is well-defined and broken down into small steps for a sequential build. It uses learning science and incorporates active recall, spaced repetition and interleaving, ensuring pupils revisit materials at various points throughout their tutoring journey.
• Every pupil is on a personalised learning journey. Those who need extra time with certain concepts get repeated exposure without slowing down the entire class.

4. Interleaving for enhanced conceptual understanding

Many traditional teaching strategies often group similar problems or topics together – sometimes referred to as “blocked practice.” However, research suggests that interleaving – or mixing different types of problems and subjects – can lead to more effective learning.

Interleaving challenges students to constantly adapt and differentiate between concepts, promoting a deeper and more flexible understanding. The principles of interleaving can be set at a curriculum level and a lesson level. Although preparing such lessons can be laborious for a human teacher or human tutor.

Interleaving is one of the cognitive processes involved in helping different students build durable knowledge.

Assessing when to interleave

An intelligent tutoring system can be trained to evaluate a student’s level and, at the appropriate point, start interleaving concepts. For example, once an element of proficiency is demonstrated in addition, subtraction, multiplication, or division, the AI tutor can then interleave problems requiring these four operations in a single session. This not only encourages students to switch contexts and apply different strategies but also far–transfer, the process of using knowledge or skills in a context notably different from where they were originally learned.

Interleaved practice tests

AI–powered interleaved practice tests are one way intelligent tutoring systems could implement interleaving. The results could help determine if the pupil has mastered the content and can progress at their own pace or if the artificial intelligence needs to revisit certain material. Subject–matter experts must set the parameters for interleaving; done too early, it can confuse novices.

What the research says about interleaving

• Rohrer (2012) shows that interleaving can improve problem–solving skills and overall retention, compared to blocked practice.
• Taylor and Rohrer (2010) argue that interleaving supports higher–order cognitive skills by forcing students to constantly re–engage with the material in new ways.

ITS example – interleaving in practice

• Each artificial intelligence lesson includes a range of questions, including a challenge section, weaving in different types (e.g. fractions, decimals, word problems) once the pupil demonstrates some proficiency.
• Before moving on to new content, Skye checks prior topics, reinforcing interleaving and ensuring pupils stay on their toes.
• Every lesson with Skye also includes examples of formative assessment, guided practice, modelling, independent practice and challenge questions to increase exposure to different types of problems and strategies, with recap opportunities as needed.

5. Encouraging problem generation

At a higher level, an effective strategy for deepening understanding is to have students generate their own problems. This generative learning approach requires pupils to combine information, connect different concepts, and apply their knowledge in creative ways.

When pupils formulate questions or design problems, they engage in higher–order thinking that promotes deeper understanding and retention.

Once a pupil demonstrates a strong level of competence in a particular mathematical principle, the best AI tutors could prompt the pupil to devise their own questions and answer them to ascertain mastery. Such active involvement enhances metacognitive skills, enabling students to monitor and manage their learning while leveraging self–explanation.

This is possibly the most challenging area of intelligent tutoring to replicate. However, by integrating problem generation into AI tutoring systems and other AI tools, developers can ensure that students are required to think hard, not just about solving problems but also about creating them.

What the research says about encouraging problem generation

• Chi et al. (1994) show that self–explanation and generative activities significantly improve learning outcomes.
• Bransford et al. (2000) discuss how problem–based learning environments can stimulate deeper cognitive processing.

ITS example – problem generation in practice

• Currently, Skye guides pupils through structured “challenge” problems. Multi–step tasks deepen understanding, encourage higher–order thinking, and help pupils apply concepts in new ways.
• The team developing Skye, the computer-based learning system, is exploring ways to enable pupils to create their own problems in future updates. The range of possible responses and the specificity of the KS2 and KS4 curricula make providing appropriate guardrails challenging at present.

6. Error correction and reflection

Effective feedback is crucial to help students learn. Beyond simply indicating whether an answer is correct or incorrect, feedback should offer insights into the nature of errors and suggest ways to improve. Reflective, immediate feedback encourages learners to evaluate their reasoning, enabling them to understand the underlying causes of mistakes.

For example, in a history lesson, if a student incorrectly identifies a significant date, an intelligent tutoring system could provide contextual information about the event, prompting the learner to reconsider and internalise the correct timeline. Not only does this address misconceptions, but it also reinforces learning through self–reflection.

Another way intelligent tutoring systems can provide reflection and metacognition is by using practice–test results to deliver a “review and reflect” session.

Artificial intelligence can praise students for the skills they have gained, but also offer support in correcting errors. This data can feed into its retrieval and spacing algorithm, further personalising the learning experience.

What the research says about error correction and reflection

• Hattie and Timperley (2007) highlight the critical role of feedback in enhancing learning outcomes, showing that timely, specific, and constructive feedback can significantly improve performance.
• Shute (2008) suggests that such feedback encourages reflection and leads to more lasting learning gains.

ITS example – feedback and reflection in practice

• Skye has been trained to identify and address specific misconceptions, and also to understand ‘correct ideas’ rather than just correct answers. 
• Skye will ask open-ended questions like “How did you get that?” or “Why do you think that’s the correct approach?” and can interpret pupils’ explanations to identify which parts they understand and which parts need to be refined. 
• Skye will work through concepts with the individual student until they’re able to confidently articulate the correct answer and the reasoning behind it.
• If a pupil gets stuck after three attempts, Skye explains the correct method in concise steps, ensuring pupils leave with clarity rather than confusion.

7. Cognitive load management

Cognitive load theory reminds us that learners have a finite capacity for processing information. Therefore, effective intelligent tutoring systems – like human tutoring – must carefully manage cognitive load.

If instructional materials are too complex or poorly structured, learners can become overwhelmed, which reduces comprehension and retention.

For instance, an intelligent tutoring system should break down complex tasks into manageable steps with clear, concise instructions. Practice problems might be presented as worked examples: the first example is completed by the AI–powered tutor, then a second example is revealed for the student to solve themselves.

Declutter the interface

Educational technology often uses overly cluttered or confusing interfaces. Learners can become distracted from the essential content and increase mental effort as they try to decipher what is being asked or which skill is being developed.

As with all good educational tools, AI tutoring lessons should show only the visuals needed for that specific learning module, removing extraneous “decorative” imagery.

Thoughtful user interface (UI) design is also crucial so that students understand how to interact with the platform and can access the powerful tools that AI provides.

What the research says about cognitive load management

• Sweller’s (1988, 1994) work on cognitive load theory offers a framework for understanding how instructional design affects learning.
• Paas, Renkl, and Sweller (2003) demonstrate that when cognitive load is effectively managed, students are more likely to engage in deep learning and transfer knowledge to new situations.

ITS example – cognitive load in practice

• Lessons are broken into single screens, each covering one short, structured step, so pupils tackle just one piece of the puzzle at a time.
• The lesson format – “I do, we do, you do,” followed by challenge questions – is consistent for each lesson, so pupils can focus on the substance, not on navigating a new structure every time.
• The shared lesson screen is intentionally simplified, with blur, highlight, and pointer tools guiding pupils’ attention to relevant parts.
• Human teachers thoroughly test and refine Skye’s prompts so that language never overloads pupils; every instruction is purposeful and easy to digest.

Effective personalisation in intelligent tutoring systems

While personalisation remains a cornerstone of effective tutoring, it is important to distinguish evidence-based approaches from popular misconceptions. Research by Pashler et al. (2008) and Coffield et al. (2004) has convincingly debunked the widespread notion of “learning styles,” finding minimal evidence that matching instruction to supposed visual, auditory, or kinaesthetic preferences improves outcomes.

Yet talk of personalisation through “learning styles” proliferates within many intelligent tutoring system platforms.

Instead, truly effective personalisation in artificial intelligence should be student centred and focus on strategies with proven impact:

• Adative teaching: Adapt to a pupil’s prior knowledge and identify learning gaps
• Adjust teaching: Adjust the pace based on demonstrated mastery rather than a predetermined schedule
• Scaffolding: Provide support targeted to specific misconceptions
• Responsive teaching: Respond to emotional cues such as frustration or engagement
• Connection: Link mathematical concepts to the pupil’s personal interests and experiences
• Timely feedback: Offer feedback specific to the pupil’s work and thought process

These evidence-based approaches respect the individual nature of learning while avoiding unsubstantiated theories.

Conclusion: the furture of ITSs

Intelligent tutoring system developers will be far better equipped to create artificial intelligence tutoring systems and other tutoring systems that genuinely enhance lifelong learning for students and effectively support teachers in UK classrooms through personalised tutoring if they adhere to the 7 principles above. We can then collectively ensure that pupils, when sitting in front of a computer assisted tutor, are genuinely improving their learning.

The potential of generative AI in education is significant, but its success depends on the thoughtful application of established pedagogical strategies. By grounding intelligent tutoring systems in these evidence–based strategies, developers can help ensure that technology acts as a valuable partner for educators, enhancing their practice rather than replacing their expertise.

As we look ahead, the most effective educational tools will be those that combine human teaching excellence with the adaptable and scalable capabilities of AI.

Evidence-based personalisation with Skye

Third Space Learning’s expert maths teachers have built the new voice-based artificial intelligence Tutor, Skye, to embody best practices in personalisation and evidence-based teaching strategies to support each pupil’s learning journey. 

Skye personalises each pupil’s journey through diagnostic assessments that identify specific knowledge gaps, adapting its pace based on demonstrated understanding during each lesson, and provides tailored support for misconceptions.

The voice-based interaction allows Skye to respond directly to pupils’ responses, while the structured yet flexible curriculum connects mathematical concepts to real-world contexts. 

This approach ensures that pupils receive genuinely adaptive instruction grounded in evidence-based practice. 

Thanks to the integration of intelligent tutoring systems with expert maths teaching, Skye provides a very low cost, schools-focused option for making effective one-to-one maths tutoring accessible to every child who needs it.

Find out more about the AI policy you’ll need in school and how schools can make the most of voice-based AI tutoring.

FAQs about intelligent tutoring systems

References

• Biggs, J. & Tang, C. (2011). Teaching for Quality Learning at University. McGraw-Hill Education.
• Bransford, J., Brown, A. L. & Cocking, R. R. (2000). How People Learn: Brain, Mind, Experience, and School. National Academy Press.
• Chi, M. T. H., Bassok, M., Lewis, M. W., Reimann, P. & Glaser, R. (1994). Self-explanations: How students study and use examples in learning to solve problems. Cognitive Science, 18(2), 145–182.
• Coffield, F., Moseley, D., Hall, E. & Ecclestone, K. (2004). Learning styles and pedagogy in post–16 learning: A systematic and critical review. Learning and Skills Research Centre.
• Darling–Hammond, L. (2010). The Flat World and Education: How America’s Commitment to Equity Will Determine Our Future. Teachers College Press.
• Hattie, J. & Timperley, H. (2007). The power of feedback. Review of Educational Research, 77(1), 81–112.
• Karpicke, J. D. & Roediger, H. L. (2008). The critical importance of retrieval for learning. Science, 319(5865), 966–968.
• Pashler, H., McDaniel, M., Rohrer, D. & Bjork, R. (2008). Learning styles: Concepts and evidence. Psychological Science in the Public Interest, 9(3), 105–119.
• Paas, F., Renkl, A. & Sweller, J. (2003). Cognitive load theory and instructional design: Recent developments. Educational Psychologist, 38(1), 1–4.
• Rohrer, D. (2012). Interleaving helps students distinguish among similar concepts. Educational Psychology Review, 24(3), 355–367.
• Shute, V. J. (2008). Focus on formative feedback. Review of Educational Research, 78(1), 153–189.
• Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257–285.
• Sweller, J. (1994). Cognitive load theory, learning difficulty, and instructional design. Learning and Instruction, 4(4), 295–312.
• Taylor, K. & Rohrer, D. (2010). The effects of interleaved practice. Applied Cognitive Psychology, 24(6), 837–848.
• Vygotsky, L. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Harvard University Press.