Written by Deepak Khemani
Once upon a time, the gurukula was the fount of education. Then came the schools. But education was available only to a privileged few — till the advent of technology.
The invention of writing allowed access to knowledge without a teacher. The printing press multiplied the written word manifold. Libraries provide access to many. Then distances shrunk to zilch with the Internet. YouTube facilitated dissemination. Online lectures proliferated. The learner became spoilt for choice.
Computer graphics and animation made concepts easier to digest. Now software can illustrate how numbers go into a balanced binary search tree. Or a visualisation of the advances that Alexander made with his army. Or the spread of a viral infection. Information dissemination has never been so enriched.
But the greatest promise of AI has been in personalised learning. Imagine a patient personal tutor a student can interact with and learn to solve problems. She poses a problem and the system generates a solution and an explanation. How can this scenario be realised? Reflect first on the need for the system to be trustworthy. Young children have complete faith in their teachers, whose word often counts against that of a parent. This should not be broken. Clearly the machine needs to have knowledge to answer the learner’s questions. The seat of knowledge is memory, which is populated by experience or by instruction.
The first thought that no doubt crosses your mind is a large language model (LLM) or a system built on top of that. So let us deal with that first. Almost everyone has experimented with ChatGPT, or Gemini, or Claude. And almost everyone has encountered a situation where the answer has been wrong or nonsensical. This is because, as Ted Chiang wrote in the New Yorker, deep inside these systems the underlying “generative AI is like a copying machine that produces a blurred copy”. The original text is often lost but a paraphrased version remains. As Subbarao Kambhampati explained in an article for The Hill, LLMs essentially do next word predictions “on steroids” so that they “can complete entire paragraphs” and a large number of parameters trained on humongous amounts of data results in “a surprising level of local coherence in the text produced, making it look like ChatGPT is logically answering the questions posed by the user”.
And yet, these models have no notion of truth or falsity. LLMs are prone to hallucinate or make up false information. A well-known example is a case from the court in New York in 2023, “in which the court sanctioned an attorney for including fake, AI-generated legal citations in a filing and explained why it was unacceptable” (Hon Ralph Artigliere (Ret), in JDSUPRA, January 22, 2025). In a recent book two computer scientists from Princeton, Arvind Narayanan and Sayash Kapoor, compare such LLMs with snake oil.
So why ever would one expose a young learner to such an untrustworthy teacher? Especially if an option of creating a trustworthy one exists.
Knowledge and reasoning go hand in hand. There are three kinds of reasoning – abduction, deduction, and induction. Machine Learning (ML) is essentially induction – generalisation from observations. Abductive reasoning, which infers causes from symptoms, is mostly guesswork. Induction is too, but it learns ‘models’ from the data which is what LLMs are. ML has tremendous applications, for example training neural networks to identify a disease from radiographs annotated by expert physicians. Such systems can often perform better than individual doctors, especially inexperienced ones. Even so, diagnosis is an abductive process with no guarantees. Doctors often rely on confirmatory tests. Such systems are not suitable for teaching, which needs to be sound, always giving correct answers. Never to mislead a young mind. A young friend asked ChatGPT a few years ago why protons are heavier than human beings and got a very coherent looking explanation.
Deduction, on the other hand, is sound. If all men are mortal, and if Socrates is a man, then it is necessarily true that Socrates is mortal. Logic and knowledge representation have long focussed on soundness. If a machine deduces an answer, then it is guaranteed to be correct. If it does a complex arithmetic calculation we can be sure that the answer is correct. Which is more than what we can say of something a LLM gives us. Practical logic systems will also have to be complete. Completeness means that if there is an answer to a problem then the reasoning algorithm will find it. An associated property is consistency. A logical reasoning system will never contradict itself.
Soundness can be achieved by carefully choosing the rules used to make inferences. Completeness depends upon how much knowledge the machine has in its memory. Unlike LLMs which digest huge amounts of text to store them in unfathomable weights in a neural network, symbolic AI systems represent this knowledge explicitly. Reasoning algorithms in Symbolic AI work on symbolic representations. This is the Good Old Fashioned AI. Explainability is an added benefit. Systems can explain how the answer is derived and why it is correct.
The bottleneck is, of course, knowledge acquisition. Where will this knowledge come from? In the heyday of symbolic AI (1970s and 80s) such knowledge was handcrafted, typically by PhD students. With the resurgence of neural networks in the last thirty years, attention shifted to machine learning. But neural network models are compressed representations that are not interpretable. On the symbolic side there have been efforts to create knowledge bases, for example WordNet from Princeton, and DBPedia, that store knowledge graphs. One must abandon the quest for creating an all knowing universal knowledge machine, and focus instead on domain and application specific ones. Therein lies an opportunity. Imagine a team of young developers building a knowledge base for high school physics, or maths, to create lesson plans and articulate problem solvers that can interact with students.
Good old fashioned coconut oil it would seem is still better than snake oil, even in the 21st century.
The writer teaches AI courses at Plaksha University
