Your Own Digital Aristotle Is Here (Nearly)
In Ancient Greece, Alexander the Great was personally tutored by Aristotle. Perhaps its our time for us to have access to such a teacher too.
With the ever-increasing integration between educational systems and technology, the possibility of simulating a human tutor has become manifest. Large language models, despite their pitfalls, have become capable of recognizing and creating vast amounts of information, cheap and on the fly. On the same day I decided to write this newsletter, the creator of Khan Academy announced in a blog post something I had been eagerly awaiting for some years: a virtual tutor embedded into learning courses to offer guided teaching and Socratic questioning. In this post, I give some background around e-learning and virtual tutors, discuss the growth and impact of AI on education, and finally explore whether GPT-4 meets the criteria for the mythical Digital Aristotle. GPT-4 is the latest and possibly best algorithm to date that can “understand”, interpret and generate human language.
A virtual tutor is a piece of computer software that can replace many complicated skills possessed by an effective human tutor. On the latter subject, we can conceive of tutors that not only allow us to learn, but that cultivate education and growth from within us. In ancient Greece, Alexander the Great was personally tutored by Aristotle. He tutored Alexander on topics like medicine, science and philosophy. He got him to read the works of literature that inspired his campaign to conquer much of the known world, including Persia, Egypt and vast regions of India. None of us (I hope) are aiming to lead armies into a world-wide conquest, but if we had a teacher as effective as Aristotle, perhaps there are no limits to what we can accomplish.
Software for educational purposes has its origins in CAI (Computer-Aided Instruction). Since then, attention has shifted towards ITS (Intelligent Tutoring Systems), which is the pairing of AI (Artificial Intelligence) with CAI. Historically with AI tutoring systems, the focus has been on expert systems i.e rule-based systems. Only recently has the field shifted towards more advanced techniques such as Machine Learning (ML). In a literature survey by Alkhatlan et al, they conclude that ‘the close marriage of ITSs with AI and psychology shows continued promise for the advancement of ITSs’.
AI and Big Data have become prominent in the past two decades. They have become increasingly used in ITS, often paired with Natural Language Processing (NLP) and Reinforcement learning (RL) techniques. The world as we know it has just been disrupted, for better or for worse, by advancements in NLP. This branch of AI focuses on the interaction between computers and human language. Specifically algorithms that can “understand”, interpret and generate human language. Most of you are probably familiar with Chatbots, which have been making all kinds of news, leveraging state-of-the art NLP.
Here are some ways in which AI-powered e-learning may be impacting education:
Improved Learning Outcomes: AI-powered e-learning can provide students with a personalized learning experience that is tailored to their unique needs, improving learning outcomes.
Enhanced Engagement: AI-powered e-learning platforms can use gamification techniques and interactive content to make learning more engaging and enjoyable for students.
Increased Access: AI-powered e-learning can make education more accessible to students in remote or underprivileged areas, providing them with the same quality of education as their peers in more developed regions.
Improved Efficiency: AI-powered e-learning can automate certain aspects of teaching and grading, freeing up teachers' time to focus on more high-value activities.
While a virtual Aristotle could make life a lot better for students, it is not straightforward to create such an entity. Before the emergence of large language models such as ChatGPT, it was unclear whether we could ever create such a thing. This is because of the highly involved process that is learning. Learning is one of those things that is both enjoyable in and of itself, while also being crucial for achieving goals and improving society. Education is something that everyone deserves, but the current state of the world doesn’t permit everyone to have a successful education. In schools, especially in marginalized communities and regions, the cohorts are too big and teachers are too few. Most students have different learning paces. What we really need is a highly-skilled teacher for every student. But this remains a pipe dream.
Teaching is a noble profession, but because of low salaries and limitations on career progression, there just aren’t enough teachers. And even if we had more, it’s unlikely that teachers could adjust to each student, to fit their particular style and pace of learning. We all know people who dropped out of education because of the lack of relevant learning. These AI-teachers simply don’t exist. Or maybe, through GPT-4, we are on the cusp of having virtual tutors take over the mantle of the modern education system.
Even if the technology were possible, market demand is what will ultimately realize AI-based tutelage. Educational technology (Edtech) is on a skyward trajectory. Accuray’s research report estimates that by 2025, the market value of e-learning could exceed $325 billion. According to a report by HolonIQ, the global e-learning market is expected to reach $404 billion by 2025, with a compound annual growth rate of 8.2% from 2020 to 2025. Additionally, the AI in education market is expected to grow from $537 million in 2018 to $3.68 billion by 2023, at a compound annual growth rate of 47%. According to a study by PwC, the contribution of AI to the worldwide economy will be $15.6 trillion by 2030. With the sudden advent of Diffusion models and large-language models, this figure could be a lot bigger. Even if we’re being pessimistic, AI is already a multi-billion dollar industry due to a seemingly implacable demand. The education sector may see a fair share of this market: unlike many other ML applications, it offers more to investors than mere profitability.
If the market demand is there, then that leaves us primarily with the problem of technology. What is needed to design a functioning Digital Aristotle? There are several companies making leaps of progress towards developing it, even excluding the ones using large language models such as GPT-4. These include Korvid AI, Squirrel AI and various language learning platforms, which have been using ML for keeping track of student proficiency. The likes of Squirrel AI also record students’ faces & patterns of attention, raising privacy concerns about AI-driven surveillance risks. IBM’s Watson has been paired with Pearson Education, using AI to give students personalized content and giving insights to their actual instructors. But none of these programmes offered are on the level of the elusive Digital Aristotle. To have a good understanding of what is needed to realize the technology, I’m going to outline key features the AI tutor needs And then we’ll discuss whether or not we are truly on the verge of a fully-realized Digital Aristotle.
To offer personalized content, we need a model that identifies the attributes that impact an individual’s learning. Let’s take a law of physics as an illustration:
———————————
Newton’s 2nd Law: force is equal to mass times acceleration.
———————————
This concept isn’t intuitive to most, but there are many paths to understanding it. Perhaps some students want graphs and diagrams- the use of imagery to aid their understanding. Other students might find the math relatively easy, but would need a physical demonstration to fully apprehend it. Effectively, you could have all of your teaching material fully labelled, to account for just about every style and format the material could possibly be in. These labelled documents could be fed into an ML model, which is a kind of AI that learns to think for itself by using lots of examples. In order for it to identify the attributes that are crucial to your learning, the model needs a reward function, a way of measuring its own success, such as the student’s test scores. After this model learns what helps and what hinders your learning, it feeds these results into a curator algorithm. Educational websites, e-books and educators on YouTube have never been more abundant. Even niche topics have so many people producing informative content around it that you can afford to be picky. When you have so many different sources, a curator algorithm becomes a truly powerful tool. Fortunately, large language models have been scraped over billions of documents from the Internet, from websites, articles and books to social media platforms or forums. These models can be given prompts, input phrases which trigger it to be have in a certain way. Something like telling the virtual tutor to use Socratic questioning to help the user learn. It can also be fine-tuned on particular documents. All that’s left is to take large language models and fine-tune them on the textbooks and materials for a particular group of students. The quality and nature of the algorithm will directly impact student learning, so choosing the right training data for the model will be crucial. The result is a chatbot rapidly picking and choosing information, taking snippets from different courses, lectures, quizzes, providing context and answers with pinpoint precision. It doesn’t have to be exclusive to online sources either. Teachers could train the model on their materials, and students could upload study notes, building up a body of knowledge, which could lead to a kind of Library of Babel, a universal platform storing every possible piece of knowledge.
The biggest barrier with this approach is that you need labels for all of this data. A lot of labels. These labels can be highly granular, for example “GCSE, Physics, 16 year old reading age, section 1, diagrams used, section 2, analogous reasoning used, section 3, mathematical formula used”. That sentence was probably exhausting to read, so you can imagine the time and energy it would take to label all that data. Thankfully, recent tools have emerged in data science to make this far more doable. There are methods that allow you to label millions of data points with just a few lines of code (such as Snorkel AI). This approach could enable people with domain knowledge, such as a teacher, to use their expertise to influence or inform the AI system. And then we have approaches like Amazon’s Mechanical Turk, where people are paid nominal wages to hand-label data (although, sadly, this is still true of even the biggest and latest AI capitalist projects). When it comes to training a useful algorithm, there are various approaches that don’r require human-crafted labels. This is a feature of Transformer-based large language models, released back in 2016. These are present in the ever-popular ChatGPT and GPT-4. These models are given sequences of text; to simplify, they learn by masking out certain parts of the sequences and then try filling them in. Repeat this process with trillions of sentences and paragraphs, and soon, perhaps you have something that could serve as the foundation for an actual Digital Aristotle.
AI requires context to determine the link between individual words, sentences and paragraphs. Previous large language models like BERT and GPT-2 were interesting, but after reading a generated sentence or two, you can clearly see that the models are regurgitating words and phrases in superficial patterns. This is still largely the case for GPT-4 (due to a concept called Stochastic Parroting), but it is good at identifying and generating relevant answers to user queries, with a narrower gap between deep knowledge and mere simulation of knowledge.
What makes GPT-4 so special, then? It can already outperform many students in a large number of tests and entrance exams. To name a few, it has received high grades from tests by ivy-league business schools, has passed exams for medical licences and also for law bar exams. This means AI already has the knowledge and ability to answer questions across a variety of domains. It can not only understand student queries in a very detailed and useful way, but it can also use visual recognition to extract information or give suggestions. All GPT-4 needs is a look at your latest homework assignment, or drafted essay, before it’s ready to fire off with guidance and suggestions to improve your grades. But it’s still limited by Stochastic Parroting, the idea that AI finds phrases or sentences that are statistically likely to be relevant to the user’s request, but may not have any underlying meaning or semantic coherence.
For AI to retain the meaning of text, there is an approach called Argument Mining, which allows us to identify what's actually being proposed in a given text. Recent focus on improving large language models has centered on “Chain-of-Thought” prompting. This involves generating outputs from a series of intermediate reasoning steps before arriving at a conclusion, reducing the harm or risk of generating meaningless or biased text. All of this is reminds me of the expert systems historically used in software-based tutor systems.
The tech is possible. The tech is in demand. The tech will help people who are already interested in learning, but what about everyone else? I believe that a Digital Aristotle could help people find their passions and aptitudes, even if they don’t consider themselves learners right now. Education is not there to make you more intelligent, but to help you realize you are already intelligent. Perhaps there could be a kind of Netflix for education, with a powerful recommendation system for courses and online teachers. There’s a product called Skillshare that markets itself in this way, but it’s not even close. With the Digital Aristotle, we don’t limit these merely to courses and books. In modern times, people eat up podcasts and short, sharp videos. This system could be a multi-modal recommendation engine. You might get a specific book recommendation based on your taste in podcasts, or get a research paper suggestion based on your taste in movies. There could be certifications and awards for course progression, such as discounts at cafes, restaurants, and for purchasing tools that can aid people in their learning.
Even after the AI tutor enriches student learning, the students will still need to write exams. A major criticism with large language models, such as ChatGPT, is that even if they can pass tests, they are only fulfilling a metric, not a target. The risk of an AI tutor maximizing a metric without actually teaching a student anything is immense. While offering this Digital Aristotle, society could also change up the structure of examinations as well. It never made sense to sit dozens of students in a big, gloomy hall, forcing them to solve problems under copious amounts of pressure, covering content that some may have covered before, while others haven’t. Standardization is important, but so is personalization. Why not let the AI tutor give each student a bespoke exam? This could ensure a fair examination, although it raises further problems and questions too, especially around the training data of the AI. At the least, it would ensure students are given a chance to demonstrate their passions and interests as well as being sufficiently challenged and penalized for lack of dedication to their education. It would distribute the grades in a more favourable way for the average student.
A privately educated student here in the UK has around a 50% chance of scoring an A or better grade in their A-Levels examinations. The question of why this happens is something explored in Outliers by Malcolm Gladwell, who makes the observations that richer kids are simply given more extracurricular learning and development opportunities than kids from lower economic backgrounds. These same lucky students will be able to go onto Ivy-league universities and have greater career options than their marginalized counterparts. So, what the Digital Aristotle offers, is a kind of private, specialized education for every student, requiring nothing more than a device with Internet access. We need to ensure this software doesn’t have its eyes glued to its profit margins, which is the case with many present AI tutor systems. Rather, AI tutors should be an open-source tool for schools around the world to adopt.
If all goes well, we’ll see the next generation as being much more well-informed and competent than us. Like Alex the Great, perhaps the next generation will reshape the fabric of the world, improving social, economic and political conditions, and sending us into the climax of the Age of Information. This sounds wonderful, perhaps even Utopian, but before we all sell our souls to this AI Mastermind, we should talk about the possible downsides of this. If we have a working AI tutor, won’t teachers lose their jobs?
The Digital Aristotle will actually free up the time and energy of the teacher, to allow them to focus on a more practical, experiential-based teaching approach. This includes more group-based activities Teachers will help students develop soft skills, such as writing techniques, critical thinking, social and teamwork skills. The upshot is that students become more prepared for the real world. I also think that this may encourage more people to become teachers in general, so they can spend their time sharing information in their own meaningful way, rather than just drilling arbitrary data points into students’ heads so that they don’t get a failing grade. The Digital Aristotle is not a replacement for teachers, but rather, a powerful tool for them to use. Going further, the presence of even a partial Digital Aristotle will be enough to make the world and our minds very different indeed. Our society will stop existing on the current processes for formal education, and instead, will morph into something that can more readily use these virtual tutoring systems.
What about students that don’t have access to computers? Won’t they be at a disadvantage? Because those students are already at a huge disadvantage, it’s time to tip the scale in their favour. E-learning is about more than just developing AI tutors. In developing countries, there is a huge demand for online learning. The six largest distance learning universities in the world are all in developing nations. The cost to bring in the IT enabling Internet access and computers in these nations is high, but we live in a time where these things are necessary for developing countries to stand against developed ones, in terms of economic power. By stimulating the quaternary sector of these regions, they have a chance to obtain much more meaningful and self-sustaining jobs than they might otherwise have had. A financial model from Michigan State University was unable to give a conclusion of the cost-benefit analysis of e-learning in developing countries. But pushing beyond this analysis, I think we should consider the snowball effect: the more IT and jobs available in these developing nations, the greater the potential for more resources and jobs is later down the line.
A serious ethical concern around using ML for tutoring systems is the fact that ML models often overfit data and don’t necessarily generalize to different demographics, outside of the ones the model has been primarily trained on. These systems may learn only from majority or over-represented populations, and provide faulty or problematic teaching or information to members of other populations. Given that there is unequal access of IT, an AI tutor could amplify the inequality of education and therefore economics. This also brings up the multi-headed hydra lurking at the heart of large language models: biased and discriminatory priors. A paper by R. Goldbach (of the Valahia University of Targoviste) contrasted the learning style of Generation Z with the learning style of other generations. There may also be divergence in learning style among students of different demographics, such as age groups.
There are psychological factors that help or hinder student learning, meaning that there are qualities of a human tutor that an AI tutor could not plausibly possess. VanLehn (of Arizona State University) discusses the ‘warm body effect’, wherein students may be able to learn from the physical presence of a person, but the paper also points out a study determining that text-mediated human tutoring produced the same learning gains as face-to-face tutoring. Another report, by Azevdeo (of University of Central Florida) concluded that emotional affect is a driver of learning, in regards to the congruency of the tutor’s facial expressions to the relevance of the content. This was evaluated using video clips of an actress, but in order to make the tutor completely automatic, it could use animated avatars to achieve the same results. Though, this may not be possible due to the ‘uncanny valley effect’, wherein synthetic faces may be missing some features that we expect to see in a real face, which can cause emotional distress for the person interacting with the face. It is clear that synthetic avatars are getting more and more human-like. More relatable, but not necessarily more realistic.
The integration of educational systems and technology has paved the way for the possibility of simulating a human tutor through virtual tutoring software. While historically, the focus has been on expert systems, recent advancements in AI and machine learning have made it possible to create more sophisticated virtual tutors. The emergence of large language models such as GPT-4 has given hope to the idea of creating a Digital Aristotle that can offer guided teaching and personalized learning experiences to every student. However, creating such an entity is not a straightforward process and requires significant research and development. Nonetheless, with the current state of the education system, virtual tutors may be the solution to providing personalized learning experiences to students who have different learning paces and styles, and who may not have access to highly-skilled teachers. In this way, the development of virtual tutors has the potential to revolutionize the education system and make education more accessible to all.
On a closing note, I’d like to share a poem by Walt Whitman:
When I heard the learn’d astronomer,
When the proofs, the figures, were ranged in columns before me,
When I was shown the charts and diagrams, to add, divide, and measure them,
When I sitting heard the astronomer where he lectured with much applause in the lecture-room,
How soon unaccountable I became tired and sick,
Till rising and gliding out I wander’d off by myself,
In the mystical moist night-air, and from time to time,
Look’d up in perfect silence at the stars.