GenAI is Hyper-Personalizing Education

Generative AI allows teachers to generate customized lesson plans and MicroSims for every student quickly.

Dan McCreary
10 min readNov 22, 2023
An image of a wise teacher surrounded by AI-powered teaching assistants. In a few years, thousands of intelligent agents will customize lesson plans and build MicroSims—image by the author and DALL-E 3.

Last week, I built my first customized generative AI application (GPTs) for generating MicroSims for Education using the new OpenAI GPTs tool. This was done with exactly zero lines of code. Although the application is far from perfect, it has already allowed me to improve my productivity 10x at building custom simulators and lesson plans for educators. As our community adds more curated examples of high-quality examples to the training set, the process will only get better.

In this article, I will review how education has been done for the last 2,000 years and the problems with the one-lecture-for-all delivery method. I will describe how modern learning management systems can use generative AI to create and match customized lessons for each student.

The summary is that education will change more in the next five years than in the last 2,000 years. But extraordinary claims demand extraordinary evidence. I give you this evidence here.

Problems With The Current Direct Instruction Model

Direct instruction has been with us for well over 2,000 years with little variation. Classroom lectures are easy to set up but hard to customize for the needs of individual students—image by the author and DALL-E 3.

Direct instruction, or “The Lecture Method,” is a teaching model where a teacher delivers a single lecture to all students in a classroom. In this approach, the teacher is the primary source of knowledge, and students are the receivers of this knowledge.

The direct instruction method is characterized by its one-way communication style, where the teacher talks at a fixed rate, and the students listen and take notes. The key element of direct instruction is a lack of feedback loops between the lecturer and the students. If you have ever taught a class on Zoom when the students have all their cameras off, you know what I mean. Physical classrooms at least allow the instructor to read the body language of the students and adjust the speed of the lecture based on their viewing of the audience's reaction.

Direct instruction is what we call our classical or traditional form of teaching, and it is the most common method used in K-12 education, colleges, and universities today. Direct instruction is often coupled with homework assignments and periodic exams that may or may not cover the topics in the lecture. Some of my college courses had only a small mid-term and a large final exam. That was the only time the instructor got feedback about their performance on sharing their knowledge. Many instructors failed, and the grades were reflected in the student's report cards.

The Roll of Feedback in The Classroom

Students use a Student Response System or “Clicker” to vote on how clear a lecture is. It is feedback but has low bandwidth, and it takes time to change the content to meet the needs of the students— image by the author and DALL-E 3

There are a few small variations from one-way knowledge transfer that involve minimal feedback and customization of content. In smaller classrooms, the students are allowed to ask questions of the instructor. The teacher can then offer clarifying knowledge or apply knowledge to a specific situation.

Some schools have tried implementing Student Response Systems (SRS) where students are given a small device called a “Clicker” that they use to give feedback to the teachers on how well they understand the content. Some schools use low-cost laptops such as Chromebooks that allow students to provide web-based feedback to the teacher during the lecture. However, these are rare exceptions today.

Other more ambitious schools are using a “flip the classroom” approach. This involves having the students watch a pre-recorded lecture and then using the classroom for questions. Teachers who use sites such as the Khan Academy have often shown good results. One key difference is that students who are struggling with the material can rewatch lectures until the concepts are well understood.

Despite all the attempts to use SRS and flipped classrooms, things have changed very little since the time of Socrates in 400 BCE. But this is all about to change!

How ChatGPT Creates Customized Lesson Plans Today

Today, tools like ChatGPT can quickly generate customized — “Destomized??” - lesson plans—image by the author and DALL-E 2. Yes, DALL-E is not perfect!

Today, less than a third of the teachers I work with are becoming aware of the power of generative AI in the classroom. Many teachers are annoyed that generative AI can help students generate text to help their writing. And there are ethical concerns about whether the writing is original work. But most teachers will accept these tools just like they accepted calculators and spell checkers in the past. Yes, they will have to change the way they teach and will not be able to make the assumption that students don’t use generative AI at home. But we can’t and shouldn't ban the use of technology for accelerating learning. Our teachers and curriculum need to adapt to the times. And we need funding and a massive volunteer effort to help them.

How does this work? Today, any teacher with internet access can use a tool like OpenAI’s ChatGPT or Anthropic Claude to generate simple lesson plans. They give it a subject area, a topic, a grade level, and other criteria they need, such as the time of the lesson and resources that the students have access to. I wrote blogs about this back in September of 2020. Every few months, the tools get more powerful. Today, the methods I outlined back in 2020 are getting better, and they have names like “Prompt Enrichment” and “RAG” (yes, a terrible name).

Although quickly generating a single lesson plan for a classroom is going to help our overworked teachers, it is not close to the solutions that are available today. Now, let’s take a look at why.

The Problems with One Lesson Plan For Everyone

Direct instruction is optimal when all students are the same. Image by the author and DALL-E 3

Direct instruction is clearly the most “convenient” way for a school to set up its instruction. You bring in people with knowledge that can create a PowerPoint deck with simple bullet points on them. The teacher hopes that all the students can learn at the same rate. They hope that they don’t lose the slower-to-learn students and that they don’t make the lecture too boring for students who learn quickly.

In many schools, there is a huge focus on making sure that no students are “Left Behind.” This is a very hot topic because it makes school administrators focus resources on the students who need the most help. This takes resources away from the talented and gifted students. And there are never enough resources in most school districts.

A bell curve showing the number of students at various learning rates. Teachers need to build lesson plans that meet the needs of the slower learning in each classroom—image by the author.

As a result, teachers are forced to build lesson plans targeting the slower learners in each classroom. And the more diverse a classroom is in terms of learning rates, the wider the bell curve. Sometimes, the students that are impacted the most are the students that have the most potential. Instead of being challenged and inspired to push themselves further, they learn that education is a slow and painful process and that schools don’t care about their needs. This is one of the greatest tragedies that generative AI can prevent.

In reality, all our students are different. We can embrace this diversity with the help of generative AI—image by the author and DALL-E 3.

When we draw the Bell curve of learning rates, it is a good first approximation of what is happening in the classroom. But in truth, placing all our students on a single linear dimension of learning rate is not a precise model of the world. In reality, some students have a better background when learning a new concept, and some can grasp math problems quickly, whereas non-native English speakers may have problems with understanding English. The truth is that every student is different, and they all learn different concepts at different rates. In the real world, there is no single single dimension for learning rates. But we don’t have 10-dimensional graphs to show how it really works. Sorry.

However, we do have tools to deal with this incredible and wonderful diversity. We have generative AI that can customize a lesson plan based on each individual person’s background and learning history. If you are familiar with generative AI, you know that if you can get this learning history into a prompt, the resulting lesson plan can be customized to the needs of the student. Now, the question in front of us is, where do we get this knowledge to enrich the prompts that we feed our large-language model?

Hyper-Personalization Workflows

A workflow of hyper-personalized lesson plans. This diagram was generated by ChatGPT using the mermaid format—image by the author and ChatGPT.

The diagram above is a high-level picture of how custom lesson plans can be generated.

On the left side of the diagram above, we see that we need to get a learning history from the Learning Management System. This could be something as simple as the results of the last quiz for the student or as complicated as an AI-generated report that puts attention on all the right elements of a student’s history, including likes, dislikes, goals, and interests. If we have too much information in the LMS, then a generative AI can be used to summarize the key facts from the learning history.

We then combine the standard lesson plan with the summary and feed it back into the large-language model such as ChatGPT. With the right prompt, the lesson plan will be tailored to the needs of the student. The result might include reading, exercises, simulations, and quizzes. It could even try to apply the lesson to a question that came up in the past from the learning history.

In the ideal world, the new lesson plan would be reviewed by the teacher for accuracy. The teachers could make changes and provide feedback that would be used to make the next lesson plan more precise.

Why Traditional LMS Systems Fail To Generate Customized Lesson Plans

From Counts and Amounts to Comparison and Similarity. Legacy knowledge representations of learning are holding back education—image by the author.

Today, most LMS systems are designed around old, non-scalable relational databases. These databases suffer from the legacy that their core data structures are driven by what could be stored on punch cards, flat files, or tables. They focus on transactions into tables and generating reports such as “What was the average score on the quiz last week?” or “What was the question that was most frequently incorrect?”

However, most LMS systems today have no tools to compare unstructured content with similar words and concepts automatically. They can’t easily generate graph representation embeddings for students, teachers, courses, content, lesson plans, questions, or assessments. They can’t even tell us what slides in a curriculum are the most similar. This forces us to duplicate content constantly. Without these embeddings, they can’t easily compare items, and they can’t make recommendations based on prior history. They are not built using modern knowledge graphs, and they can’t easily leverage the advances in generative AI.

In the past, comparing content and finding similar content was not hard. LMS systems had very little content, and simple metadata tags and keyword searches were all you needed. But generative AI is going to change all that by its ability to generate vast amounts of high-quality content.

The Coming Flood of Low-Cost, Easy-to-Customize MicroSims

The era of boring drill-and-kill content is over. In its place, generative AI is building easy-to-customize interactive MicroSims coupled with storytelling—image by the author and DALL-E 3.

For the past several months, I have been working with a visionary group of technologists and educators who are teaching teachers (and students) how to build MicroSims. This was a term coined by my colleague Valerie Lockhart this year. At the core of MicroSims is a design process that partners educators and students with generative AI tools to generate simulations that are useful to a classroom. Central to this process is the fact that generative AI does not just generate the first pass of a MicroSim. You can continually modify your MicroSims or other MicroSims.

Although the concept of using ChatGPT-4 to generate code is not new, what is new is how it is coupled with frameworks like p5.js. P5 is a modern variant of the powerful Processing language that has been around for over 20 years. There are tens of thousands of open-source simulations already available. And it seems as though ChatGPT has been trained on them all! Just a simple prompt of “generate a p5.js sketch of a ball bouncing” is all you need to get started.

What is missing today is GPTs fine-tuned just to generate high-quality user interfaces complete with sliders that also have labels and values. We are working on that project now, and we look forward to anyone who would like to help us create high-quality curated samples.

After that, we will proceed to store these prompt/result pairs in a knowledge graph with a concept index so prompts can be enriched with similar MicroSims. We only need some volunteers and a few million dollars to build a proof-of-concept. Let us know if you have ideas of who might fund this.


In order to help us move away from the direct instruction model toward hyper-customization, we need to break from the past. We need modern learning management systems that are tightly coupled with knowledge graphs and integrate generative AI. We need vast amounts of easy-to-customize MicroSims to enable teachers to get away from the drudgery of creating lesson plans. New LMS graphs and MicroSims can focus on helping facilitate the movement to project-based learning and getting students to build things together. That is when our educational systems will really start to take off!



Dan McCreary

Distinguished Engineer that loves knowledge graphs, AI, and Systems Thinking. Fan of STEM, microcontrollers, robotics, PKGs, and the AI Racing League.