#3 Humans v/s AI: The act of reasoning

This post is part of a series on the similarities and differences between natural and artificial intelligence. You can find the first introductory post here, which includes a description of all the topics that I’ll be covering: https://ishaan-b.medium.com/the-nature-of-intelligence-in-man-and-machine-a-series-c9b6c8a5e2a6

If you’re excited by this subject, feel free to follow my Medium profile here to stay up to date with new posts: https://ishaan-b.medium.com/subscribe

— —

In my previous two posts, I defined and explained artificial intelligence and natural intelligence, respectively. In today’s post, I compare the similarities (and differences) between how each of them works. The previous posts provide useful context and background for the arguments made here, so you can refer back to them in case any concepts seem foreign.

The act of reasoning is the first lens through which we will compare artificial and natural intelligence. Keeping in mind the outcome-focused view of intelligence that we established, we will see that the key bridge between natural intelligence and AI lies in their shared ability to perform both deductive and inductive reasoning.

AI is capable of deduction and induction

Deductive reasoning happens in the left-side of the human brain and works through linear information processing, i.e. by using rules such as propositional logic statements that specify how to explore the world, and following those rules to a logically consistent conclusion. Symbolic logic is the mechanics of this type of reasoning. This is essentially what happens in all knowledge-based or planning agents, the first form of AI that was created by us humans. One example is a simple tic-tac-toe computer opponent, which follows a set of directions to choose the optimum square. Another example is a medical ‘expert system’, which analyzes a set of symptoms to match them to a particular diagnosis, based on a mapping of symptoms-to-illnesses that is contained within the AI agent. In both these examples, we know exactly how the AI agent arrives at its decision, because we know the rules it follows and the knowledge it contains. In many ways, it is nothing more than an algorithm, albeit an algorithm that executes itself.

Inductive reasoning happens in the right side of the human brain and uses parallel information processing. Different pieces of data are processed at the same time to arrive at some universal principle or truth that governs them, which can then be used to make predictions about other, related phenomena. This is what machine learning also accomplishes. Deep-learning networks are multi-layered with millions of parameters, thus enabling parallel information processing. The more data that is made available to a machine learning algorithm, the better it can understand the patterns at play, and the more accurate its predictions become.

Sir Francis Bacon, the 17th century English philosopher and statesman, investigated the nature of inductive reasoning in his work Novum Organum. He designed an experiment to identify the inductive method of hypothesis creation; his process involved systematic capture of observations in the form of data, which could be used to arrive at hypotheses that explained the data. Bacon’s method fell short in describing how a hypothesis is created if there is no direct logical connection to the data. Nevertheless, his experiment highlights another key procedural similarity between inductive reasoning and machine learning: the use of structured data to arrive at general hypotheses. The next step in this process is spontaneous insight, which was the missing link in Bacon’s experiment.

We know surprisingly little about how spontaneous insight occurs. The prevailing theories posit that it is conducted by the subconscious mind. Many psychologists have written on this subject, such as Wolfgang Kohler who investigated problem-solving in great apes and observed subconscious moments of sudden clarity. This leads us to another key similarity: we don’t know how machine learning arrives at its outcomes either. In simpler versions of supervised machine learning, the models are trained on labeled datasets, so humans do have an understanding of the initial patterns that are provided (e.g., certain words are labeled 'spam', so a spam AI agent learns to label all emails containing those words as spam). But in larger deep-learning models, with multiple layers and millions of parameters, we just throw in as much data as we can and achieve remarkable results. We have no idea how the models arrive at those results. Large Language Models (LLMs) are a prime example, where we know they predict words/sentences based on probability of those words occurring together (called ‘token-prediction’), based on what they learned from their training data. But we do not understand how LLMs produce consistent and accurate information and insight. Moreover, we don't even really understand how they learn grammar and syntax. The rules of English are not provided to LLMs; they understand how to communicate in English based on patterns in the text data they consume.

The example comparing Deep Blue and AlphaGo in my previous post exemplifies this further. Deep Blue was primarily a logic-based AI; because of its Bayesian structure, its programmers could audit the computer’s decision-making afterward and determine, in retrospect, why it had chosen to act a certain way. Deep Blue was an example of explainable AI, so its decisions were transparent and later easily understood by designers. AlphaGo was built on machine learning, and learned move patterns through repeated play. Patterns that we cannot clearly identify and explain.

I like to think of inductive reasoning as the abstraction of essence. Machine learning does something like that, too. A philosopher may argue that, when ML algorithms reach into their vast troves of input data to make predictions about the future, what they’re really doing is figuring out some basic underlying truth about the phenomena being described. Where does this truth come from? I’ll discuss this further in my future post on Big Data.

So we’ve seen that certain forms of AI are able to make similar, rational decisions as humans. Unfortunately, the resemblance between the reasoning capabilities of AI and natural intelligence seems to end there.

Differences in the way intelligence is deployed

One key difference between AI and natural intelligence that is immediately obvious is ‘cognitive imperative’. Human intelligence contains within it an innate passion to understand the world, a virtual compulsion to explain that which exists around us. Just as intelligence arose through evolution, this cognitive imperative is evolutionary in nature, for without it we would fail to identify the important aspects of the world that we need to adapt towards. This is linked to meta-awareness and does not exist in artificial intelligence by definition.

Secondly, we are able to understand other humans, their emotions and feelings, in ways that cannot accurately be described by a set of rules. It often differs by person and by context. This is very difficult to replicate in artificial intelligence. The closest we have come is through logical agents by attempting to define characteristics of various emotions (e.g. flushed cheeks, rapid heartbeat etc.), though this is neither accurate nor comprehensive. Perhaps this can be solved in the future using big data if we collect enough information on different instances of human behavior, then give it to a machine learning algorithm with access to video, audio and other ‘sensory mechanisms’ and allow it to run wild. The underlying assumption in that theory is that there are patterns we can read in human behavior, as opposed to an innate skill to understand other members of our race.

Lastly, one vital difference that’s important to highlight is the relationship between deductive and inductive reasoning themselves. The strongest AI prediction engines that we have built, which use deep-learning neural network models, are not capable of incorporating deductive reasoning. They can either solve, or predict, but cannot do both. In humans, on the other hand, deductive and inductive reasoning go hand-in-hand.

Recall the example of Gauss’s big math breakthrough. The new theorem he devised was a big leap that came about through inductive reasoning. But a theorem is not considered complete until it is proven, which was done through deductive reasoning. It was only accepted once there was a clear logical explanation (i.e. mathematical proof) of why it is true. While inductive reasoning helps discover a binding principle, deductive reasoning is necessary to test, confirm, understand, and interpret it. A machine learning model, unlike a mathematician, cannot explain how to get to an answer. And we don’t know enough about what’s going on inside it to explain it ourselves. All we can do is take the output it throws out and run with it. That’s why machine learning is still not used in scenarios where correctness is of paramount importance, such as the field of medicine.

It is interesting, therefore, that LLM output can be vastly improved by using chain-of-thought prompting. That is basically a way of creating prompts that tells the chatbot how to get the answer. For example, instead of asking a generative-AI chatbot to “write an essay on this topic”, you type “pretend you’re ___, an expert in ___ field; using ___ documents as a reference, provide an answer to ___; use the answer from your previous response to write an essay on ___”. Chain-of-thought prompting was found to elicit reasoning-like capabilities in LLMs, especially for math related problems which they are notoriously bad at. But the LLM does not suddenly learn how to perform deductive reasoning. It is not executing your rules in logical steps. It is simply producing output in discrete, individual steps, then using those outputs as a new prompt. This reduces the margin for error and misunderstanding at each step. The ‘reasoning’ that you have been able to induce in the LLM is not intrinsic.

Ever since I first learned about machine learning, I realized that it would be a major breakthrough when someone successfully incorporates elements of symbolic logic into deep-learning algorithms. At the moment this is simply an area of academic research, and that too for some specific use cases only.

So if artificial intelligence displays reasoning abilities… does it also show elements of conditioning?

I discussed the relationship between natural intelligence and biological conditioning in my previous post. Despite AI possessing the pre-requisites to conditioning, I do not believe that AI displays any elements of true conditioning.

Biological organisms use their powers of reasoning to understand facts about the world and adjust their behavior accordingly. Conditioning entails adapting to your surroundings.

AI agents are created with specific tasks in mind. They are yet to be ‘general-purpose’, unlike natural intelligence. Since AI is intended for specific situations, it faces no need to adapt to other situations. Therein lies another key difference between natural and artificial intelligence.

With that being said, it is interesting to note the emergent properties that we observe in deep-learning models such as LLMs. From the Wikipedia page about ‘emergence’: In philosophy, systems theory, science, and art, emergence occurs when a complex entity has properties or behaviors that its parts do not have on their own, and emerge only when they interact in a wider whole. Emergence is rooted in the idea that the behavior of large and complex combinations of elementary particles cannot simply be explained by extrapolating the properties of a few particles. Instead, at each level of complexity entirely new properties appear. Nothing exemplifies emergence more than life itself. With LLMs, we have a clear example of this in their ability to perform complex reasoning. The simple act of ‘token-prediction’, i.e. suggesting the best combination of words, does not explain their ability to reason and explain complex issues/problems. Clearly some greater properties emerge in the LLM once we provide it with some critical threshold of training data. I will be writing a post about Big Data where I’ll explore this further, along with some other fundamental questions about information theory and evolution.