Kant, Hume and AI
Foundational concepts and the foundational structures of human thought
Immanuel Kant is a pivotal figure in the history of philosophy. He was not, however, a precocious genius who pioneered new ideas at a young age. Instead, reading the work of David Hume famously (and in his words) ‘interrupted his dogmatic slumber’ and provoked him to rethink a lot of philosophical assumptions.1 His approach, and his reaction to Hume, helps illuminate a number of questions I've been exploring here, including the differences between human thinking and artificial intelligence. His analysis of foundational human concepts and the structures of human thought is particularly relevant. This part of Kant's philosophy is often put to one side today, but the idea, if not all the details, is worth exploring.
Causation? Or is there only correlation?
Kant's approach is easier to understand in contrast to Hume's arguments. Hume assumed that all human concepts and thinking was derived from experience and his approach to philosophy was then to interrogate traditional (philosophical) concepts to see if they stood up to close scrutiny - and often he argued that they didn't. One typical example of this approach was the way he analysed the relationship of 'cause and effect'.For Hume:
When we look about us towards external objects, and consider the operation of causes, we are never able, in a single instance, to discover any power or necessary connexion; any quality, which binds the effect to the cause, and renders the one an infallible consequence of the other. We only find, that the one does actually, in fact, follow the other........
Consequently, there is not, in any single, particular instance of cause and effect, any thing which can suggest the idea of power or necessary connexion.2
In other words, we can only ever see one thing happen and then another thing after it. We never see something causing something else to happen as we can never see that special thing that is the actual causation. I might throw an object at a window and the glass breaks, but - logically - we have no way we can see or conclude with complete certainty that the object caused the glass to break. It may have just happened exactly when the object got to the window. As a result, Hume argued that when we talk or think about causation what we really mean is that there is a 'constant conjunction' of events - whenever one things happens, the second always occurs afterwards. To rephrase this point in more modern terms: what we refer to as causation is really repeated or necessary correlation.3
In response to Hume's philosophy, Kant advocated for what he called a 'transcendental' approach to philosophy. To quote (in translation):
Human reason, in one sphere of its cognition, is called upon to consider questions, which it cannot decline, as they are presented by its own nature, but which it cannot answer, as they transcend every faculty of the mind.4
Rather unusually for a famous philosopher, Kant's position was that human thought and experience could not be completely understood. Instead, he argued that there are foundational concepts that structure how we experience and understand the world but we cannot explain or transcend them. As his terminology and writing is notoriously difficult, I will use causation as an example of a Kantian argument rather than explain his actual stated positions.
Kant took Hume's conclusion that we can only ever experience causation as one thing happening after another seriously. At the same time, it seems impossible for us humans to make sense of the world without genuine causation existing. Kant concluded therefore that the concept of causation (amongst a range of others) was a foundational structure of the human mind, or "a pure concept of the understanding". In other words, as humans, we are hardwired to see cause and effect in the world and cannot understand the world without it.
Where this concept comes from, or whether causation truly exists in the world, is something that transcends our rational abilities to decide with certainty. This hardwiring of a concept or way of seeing the world might be due to evolutionary adaptation, or how we were made by a deity, or simply chance, but it is a fact about humans rather than necessarily a fact about the world. But we do, and should, live as if causation is real. I would add that given it works so well in explaining the world it is likely to be real, but we cannot prove this, or the opposite, conclusively.5
It is worth noting that Kant's commitment to a truly transcendental approach to philosophy was limited. Based on his insight that humans experience the world in a particular way that relies on particular concepts, Kant codified a system of twelve Categories, in four neat groups of three - one of which was cause and effect. For Kant, this system of Categories indisputably described the fundamental structure of Human Understanding, and all other human concepts, in combination with experiences, are derivable from these twelve.
Hardwired concepts
Kant's system, in which there are exactly twelve foundational concepts, or Categories, seems too neat to be true. It also assumes that we can know the structure of the human understanding with a certainty that is at odds with Kant's own transcendental approach. However, the core insight that there are foundational concepts, or perhaps predispositions to experience the world in certain ways, hardwired into humans both lines up with much of our experience and helps us make sense of the structure of our knowledge.
For example, I have previously argued that all of mathematics is based on a couple of core concepts: (i) the concept of 'one' and (ii) that putting 'one thing' and 'another thing' together makes a different number of things (i.e. counting). The first of these appears in Kant's table as 'Unity' and the second is arguably encompassed in the idea of 'Plurality'. What is intriguing is that it is very hard to identify where the concept of 'one' comes from as it is instinctive for humans to identify an object as a singular thing that is not the same as things around it. In other words, we are hardwired to identify things as singular objects, or in Kantian terms, unities.
There is also good evidence that we intuitively understand the world in terms of cause and effect. Pediatricians tell us that babies start understanding cause and effect from the age of 3 or 4 months - which suggests an innate tendency rather than something that is taught by others or from experience. And all of us tend to see causation in the world more often than it occurs, hence the common refrain not to confuse correlation with causation. This also suggests that our minds are structured to jump to identifying cause and effect, rather than having learnt it from painful experience.
AI works differently
As noted recently, comparing human thinking to artificial intelligence is a useful way to better understand both. And this point, that humans have certain underived, foundational concepts or predispositions that structure how we experience and understand the world is a significant point of difference.
Last year I argued there is a foundational difference between humans and AI in how they identify objects or patterns. Humans have an innate skill at identifying things as a discrete object and we identify objects by first identifying something as an object and then classifying it. AI systems, by contrast, need to identify an object as something that is already in their training set or database, before they will 'see' that there is an object there. AI systems literally don't see obvious objects (to us) if the type of object wasn't in their training data.
This is another argument that humans are hardwired to experience the world as a series of discrete, singular objects. Or, to use Kant's terms, Unity is a Category, or pure concept of Understanding, that structures our experiences. Importantly, we do not know where this comes from or how to define it and so AI lacks this fundamental concept, structure or skill.
Causation provides another relevant contrast. For AI, or any computer program, I would argue that they can only ever experience what we call causation in an entirely Humean way - the constant conjunction or correlation of different input data. The concept of a cause over and above the logic of ‘if A happens then B happens’ is important to human thought but is absent from the logic of programming. While Hume would have been happy with this as a genuine account of causation, his account has failed to satisfy many philosophers.
This means we have another reason why AI is a very different type of intelligence to human intelligence. It also explains how humans have surprising abilities to do certain things that are really difficult to program computers to do. We have innate skills, or hardwired concepts, that shape how we experience the world in important or useful ways but which we cannot (easily) explain or define. And it is yet another reminder that we don't know as much as we think we do!
The full quote is in the Introduction to his Prolegomena to Any Future Metaphysics: “I openly confess, the suggestion of David Hume was the very thing, which many years ago first interrupted my dogmatic slumber, and gave my investigations in the field of speculative philosophy quite a new direction.” Full translation available online at https://www.gutenberg.org/cache/epub/52821/pg52821-images.html
David Hume, An Enquiry Concerning Human Understanding, Section 50. Full text available online at https://www.gutenberg.org/cache/epub/9662/pg9662-images.html
This is a simplified account that tackles one core point. For a more detailed explanation of Hume’s arguments, see https://iep.utm.edu/hume-causation/
This is the opening line of his Introduction to Kant’s Critique of Pure Reason. Full translation available online at https://www.gutenberg.org/cache/epub/4280/pg4280-images.html
Alongside concepts like causation, Kant famously argued that space and time are also structures of human experience as we cannot make sense of the universe without them. This prompted his famous distinction between the world as it appears to us (Phenomena) and the world as it is in itself (Noumena). Interestingly, String Theory in modern theoretical physics supports a Kantian view here as it has posited 8 - 10 spatial dimensions of which humans can only experience 3.
There is a long and deep debate about whether Kant believed that things like space, time and causation don’t really exist or whether his position was that we cannot know if they do exist. The outcome is not relevant to the argument here as the important point is the way we humans are structured or hardwired.
Great article, Mate.