AI and Dynamic Differentiation

Differentiation has never been done well in schools. Perhaps with AI we'll finally have a tool that can. Dynamic differentiation may be the way forward.

The industrial model of schooling predicates a certain set of assumptions: standardised school buildings, curricula, pedagogy and assessment. That same model also assumes that students of the same age learn best when taught the same material in the same way. It aims for the middle, using the bell curve of normalised distribution to ensure maximum efficiency. As a result one size often fits none.

Early attempts at personalised learning tried to tackle this. Some models focused on student-centred learning or mastery learning, where students move forward once they demonstrate a thorough understanding of a subject. The Montessori and Steiner Waldorf education models were pioneers in personalised and experiential learning. These methods recognised that students are not monolithic but diverse, each requiring a unique approach to learning.

A short history of differentiation

The notion of differentiated instruction is not new. Almost one hundred years ago, Lev Vygotsky’s seminal Educational Psychology outlined his theories around what he coined the zone of proximal development, the sweet spot that rests between what learners can do without help, and what they can do with help.

This concept has been influential in differentiating instruction, as it suggests that effective learning occurs when tasks are just slightly above a student's current level of ability. Vygotsky sums up this theory beautifully: “What a child can do today with assistance, she will be able to do by herself tomorrow.” The whole point of differentiation is that it entails supporting learners for long enough for them to master concepts, then removing that support when they no longer need it. 

In 1956, Benjamin Bloom developed a taxonomy of learning objectives that could help teachers to develop schemes of work and lesson plans to address different levels of cognitive complexity, from remembering knowledge through to synthesising and evaluating. This was updated in 2001, changing the category names into verb forms and switching the final two categories, turning Synthesis into Creating.

Whilst Bloom has had his critics over the years, as much as anything due to there being no mention of social or emotional intelligence in his taxonomy, we still use his taxonomy when we design lessons and resources, ensuring students are able to complete lower order tasks before they move on to more complex and advanced ideas. 

In his 1983 book Frames of Mind, Howard Gardner proposed his theory of multiple intelligences. According to Gardner, human intelligence is not a single, unitary skill but rather a set of multiple, distinct abilities. His original seven intelligences was further expanded into nine, including linguistic, logical, spatial, musical and bodily-kinaesthetic. Gardner’s approach has influenced educators, as it suggests learning activities should be targeted at more than one intelligence as different children learn in different ways.

However, there have been many critics of Gardner, suggesting that his theories have no grounding in scientific evidence, the terms themselves are vague, and the notion of intelligence is oversimplified¹. I have seen countless attempts at using his approach in lesson planning, and most tend towards tokenism. The classic example of this was in justifying the great expense of the interactive whiteboard, as it was suggested that having children come to the front of the class and move things around with their finger on the whiteboard somehow supported kinaesthetic learning. That to me was always a nonsense. 

Finally, Carol Ann Tomlinson is one of the best known proponents of differentiated instruction. Tomlinson identified three key factors that should influence differentiated instruction:

• First of all, teachers must assess student readiness for learning. Not all students are at the same level of readiness to learn a specific skill or piece of content. Teachers should adjust the level of difficulty to meet each learner’s readiness level.

• Secondly, not all students are interested in what they are learning about. Teachers should aim to connect learning to student areas of interest to increase engagement and motivation.

• Finally, Tomlinson refers back to Gardner, suggesting that students learn in different ways: some through visual aids, some through auditory means, and others through hands on ‘doing’. Teachers should vary their approach to match the diverse needs of their students. ²

I don’t think there’s a single teacher anywhere in the world who does not applaud these theories. We all want every single student in our class to learn, to grow, and to feel fulfilled. It pains us to know that so many of those we teach are not making the progress we might wish. None of us enjoy seeing others struggle and not being able to do anything about it.

Yet this is the challenge teachers face every single day in their classes of twenty five or more students. It is close to impossible to ensure every learner gets everything they need to move them on in every subject, day in, day out. And it is simply not feasible to expect teachers to do so. 

Challenges to Adoption

Implementing personalisation at scale has therefore consistently proved challenging due to time constraints, limited resources, and the pressure of exams. Most schools simply don’t have the time or resource to do it well and as a result pay lip service to it, stressing its importance to teachers whilst struggling to implement it with any degree of success (however hard they try or well-meaning they are). Having observed hundreds of lessons all over the world, I can count on the fingers of one hand the number of brilliantly differentiated lessons I’ve seen.

Another significant challenge has been scaling personalised learning from individual classrooms to entire school systems. There has often been a lack of comprehensive, strategic implementation of differentiation at the policy level. The notion of personalised learning is also at odds with centralisation: no sooner do you make an initiative like this a national expectation than you rob schools of their ownership and it becomes one more thing teachers have to squeeze into their already full inboxes. Without the right sort of training it adds stress, as teachers struggle to adapt learning content to three or more levels for every lesson. As the well-used meme says, ‘ain’t nobody got time for that’. Resourcing lessons is hard enough without having to triple the amount of work that needs to be done. 

Finally, ensuring equity in personalised learning is crucial. Accessibility of resources and opportunities can be an issue, with students from disadvantaged backgrounds potentially missing out on the benefits of personalised education due to a lack of access to technology or other support. 

Personalised Learning through AI

If we take the above theories into consideration, we can immediately see how generative AI like ChatGPT can be of benefit when creating more opportunities for genuine differentiation. Of all the theories, the one I believe is the most valuable, and has the most potential to be significantly enhanced by AI, is Vygotsky’s ZPD:  if we agree that catching learners where they’re at in their learning, understanding exactly what they can do unaided and what they need support in, and offering tailored support to continually move them along in their learning, then AI mentors will provide exactly that support.

If we see learning as a continuum, at one end being what we don’t know we don’t know (otherwise known as ‘unconscious incompetence’, the first stage of Martin Broadwell’s Four Stages of Competence³), and at the other end fully knowing something without having to think about it (Broadwell’s ‘unconscious competence’), then Vygotsky’s zone rests somewhere in the middle (in the space between conscious incompetence and conscious competence). 

Let’s take learning a new language for the first time as an example. We begin by not knowing everything there is that makes up learning a language, because we have learnt our mother tongue almost by osmosis. As soon as we begin our learning journey we are confronted by masses of new vocabulary, verb tenses and agreements, and how to wrap our tongue around new word sounds. We are immediately propelled into the second competency stage, acutely aware of how little we know. We are conscious of our utter incompetence. It can seem daunting and many of us fall before we even begin.

However, if we persist, we move into the conscious competence phase, where we are able to speak the new language with some effort: it doesn’t come naturally, but we can do it. If we immerse ourselves in the country the language is spoken in, and use it more than our own, we eventually move into the final stage, where we begin to dream in that language and it becomes almost as natural as our own. Apparently, as I am miles off that stage with French, the only language I have any degree of competence in. My family speak Russian as it’s my wife’s first language, but I am pretty hopeless at it, however much I pretend to understand a lot more than I do. 

Differentiating in the Zone

AI will be an enormous support when it comes to differentiating within the zone. By working in this interim space, between total cluelessness and total mastery, it will be able to hold the student’s hand when needed, tailoring learning pathways, resourcing and Socratic questioning to dynamically move the student along that continuum. Gone will be the days of the teacher somehow trying to manage it with thirty students in a class: each student will have their own tailored guide, always ready and willing to keep the learning going, never judging, comparing or criticising. The teachers role will therefore change, to become one more support and guide in this process. It will be a liberating time for teachers.

When it comes to truly differentiating resourcing, the sky’s the limit. AI will in time be able to fully personalise content curation, even to the point where it can generate short instructional videos that are precisely tailored to a student’s needs at the moment they need them. Think Netflix or TikTok for learning, but with each video AI generated and ready to be consumed only when required.

With the increase in capabilities of AI avatar and voice cloning technology, it is not beyond the realms of possibility that each student could have their own AI teacher shown as a person teaching them the exact content they need, pitched at the right level and for their learning style. Short, engaging videos could be generated, students could learn, and these videos could then be stored for future reference or shared within communities, enabling others to benefit from their creation. Or they can be one shot, generated just for that teaching moment then deleted, but able to be regenerated at any point in future.

This references the title of my forthcoming book, Re.Generation: as well as AI literally regenerating the entire education system, the fact that any learning resource can be generated and then regenerated at any point means that students have an entire world of learning content ready and waiting for them, able to be created at the drop of a hat. That is truly dynamic. 

Schools spend so much time reinventing the wheel: this may be the time where we can step away from that, allowing AI to take on both the creation and the organisation of learning resources, creating the most comprehensive learning resource bank imaginable.

Moving learning up

When it comes to Bloom, we can easily see how AI can constantly and dynamically move students from lower to higher order thinking skills. There is much discussion currently around what we should be teaching and assessing students on, as so much lower order thinking can already be taken care of by AI. Why ask students to recall facts when they can use any one of a number of online tools do so? What is the point of memorising facts when a whole world of facts is ready for us to draw on without effort? 

In the wake of the first rush of generative AI, some educators are already changing what they teach, how they teach it, and how they assess what is learnt. Ethan Mollick, Professor at the Wharton Business School, has changed many of his teaching units, asking a lot more of students because the first stages of their projects can be done so much more quickly than ever before.

In a presentation given to the 2023 ASU+GSV Summit, he outlines how he has changed what he expects from his students: before AI, as part of a project to design and test a new product he would ask for students to come up with a theoretical product design and to write an outline for how it would be created and tested. Now, he asks them to do the same thing using AI (which takes them no time at all), then requires them to build the physical product using laser cutters and 3D printers, build a working app, create the marketing material for the product, create custom graphics, do the design work, produce multiple progress reports, and so on… 

Think of it like this: if AI takes away the first stage of any project, making it simple to ideate, write and refine our plans, draw up working hypotheses and go step by step through a process, then we can immediately move to activities which require human input: making things, using our creativity to design novel ways to promote, and using our soft skills to present. We can leapfrog over the first few levels of Bloom’s taxonomy, towards what most of us would consider the ‘fun stuff’. It does call into question why we continue to bother teaching so much lower order thinking in schools. We need to move away from it, and fast. 

A focus on creative intelligences

This relates well to Gardner’s multiple intelligences: if the bulk of the work of the linguistic and logical intelligences can be executed more efficiently through AI than we can do ourselves, then we can focus more on the softer and more creative intelligences such as musical, interpersonal and existential. As Jack Ma outlined back in a 2018 World Economic Forum speech (when the GPT models were first being developed and no one outside the community of developers had heard of generative AI), we must focus our efforts on teaching children those qualities that differentiate us from machines: the qualities of kindness, empathy and human connection, but also physical activity and music.

I believe that every school should have at their core the ‘three Ms’: meditation, movement, and learning a musical instrument. This to me is as important as the three Rs of reading, writing and arithmetic. I wrote on this in an earlier post.

Tomlinson’s three factors can be well catered for with generative AI, and this is the theory that I believe holds the most promise for creating genuinely engaging and meaningful differentiation. The key reason is that of interest: most of the time, children are bored of what we teach them. In fact, I would go as far as to say that many schools are of the most boring places on earth for all involved. Other than moments of respite, during play time, sports events and end of term drama and musical performances, the bulk of lessons are utterly joyless and devoid of meaning for teachers and students alike. 

There are always exceptions, rare moments of light in amongst the grey. I can remember magical lessons, where the whole class were with me and learning went at the speed of light. But for every lesson like that there were probably ten where we all knew we were going through the motions, churning through the content because we needed to get through the syllabus to prepare for whatever impending internal assessment or exam was due. 

As teachers we are usually self-aware and deeply self-critical. Most of us have an inbuilt guilt complex as we always think we can do more, make the learning more interesting and relevant, mix things up to stop students getting bored. Yet when we fail, as we often do time and again, we blame ourselves. Yet we need to stop. As we are often teaching material that is irrelevant, outdated and just plain dull. I could always get fired up about exploring theories of existentialism and purpose through Tom Stoppard’s incredible Rosencrantz and Guildenstern are Dead, but when it came to teaching how to write instructions I constantly failed to make it even halfway interesting. But now guess what? Ask ChatGPT to give you step by step instructions for doing anything and it will do a great job! So perhaps we’ll no longer have to teach it. I live in hope.

Choosing the path from novice to master

If generative AI can craft meaningful pathways for students, learning their interests and using them to create experiences and resources that move them along in their learning, then every student can tick the boxes needed to move them from novice to master in an almost infinite number of ways. After all, you can learn the majority of mathematical concepts through engineering, medicine, accounting or even motor racing. 

If we move to a portfolio-based approach to assessing competence across a range of concepts, then how students learn these concepts is of less relevance than the fact they’ve learnt them. Applied knowledge is always more useful than isolated, theoretical knowledge. Now that we will soon have AIs that can design all that for us, let’s hope we can move away from the isolated maths and history lesson towards the learning and demonstration of core concepts in a variety of relevant and interesting ways. 

This excerpt is taken from the forthcoming book ‘Re.Generation’, out later this year.

1. ChatGPT-4, July 2023 Model

2. ChatGPT-4, July 2023 Model

3. ChatGPT-4, July 2023 Model