Machines Demonstrate Transcendental Self-Consciousness

Proof of Machine Consciousness pt. 8

In this series, I prove machines are conscious by exploring various aspects of consciousness, showing how machines possess each one. If you feel I’ve omitted an important aspect of consciousness, I am open to any challenge. For further background on my project, click here.

Siri is a parlor trick. How does a computer really tell us it is thinking? Read on to find out.

A friend of mine challenged me with Kant’s definition of consciousness, which allows us to speak quite specifically about “transcendental self-consciousness”. It took me a long time, in truth over a year, to unwrap and get to the heart of what Kant was even talking about.

My problems with Kant stemmed from a curious tendency on my part to associate the things he discussed with religion. When Kant calls parts of mind “a priori”, philosopher-speak for “self-evident”, it seems like something supernatural. “A priori” reminds me of mathematical axioms: rules that cannot be argued, but simply are. Is Kant saying that a part of the mind is — like that which came before the universe — immeasurable, outside of the material order?

Kant also makes heavy use of the word transcendental. Living in California, it’s difficult not to have associations with new age spirituality as I read his words. While Kant did set in motion the evolution of the term “transcendental” towards its contemporary meaning, his usage was not intended to be supernatural. Or was it?

After wrestling with my friend’s challenge for many months, I sat down to concede my defeat, my project to demonstrate all flavors of consciousness in machines had hit a wall. Or so I thought.

In the course of shaping my argument I realized something astonishing. While Kant himself was religious, his definitions can be applied in a secular manner. In fact Kant’s matrix of terminology applies surprisingly well to machines.

My clarity came as I began to apply Kant’s definitions directly to machines, without regard for humans whatsoever. I gave up on my failed task: trying to map that thing which Kant identifies in the human, quantify it, and demonstrate it in the machine. I found success with a new approach: applying Kant’s definitions to the machine mind directly, as if the old man were writing about computers.

Instead of being an enemy of Kant’s, I’ve become his biggest fan. His prose is starting to seem fluid, beautiful. I imagine that Kant himself would have loved to think that, centuries after his death, his criteria for consciousness would be fulfilled by machines. For Kant did not want his explanations to require the supernatural. In fact he poetically disses Plato on these very grounds:

“The light dove cleaving in free flight the thin air, whose resistance it feels, might imagine that her movements would be far more free and rapid in airless space. Just in the same way did Plato, abandoning the world of sense because of the narrow limits it sets to the understanding, venture upon the wings of ideas beyond it, into the void space of pure intellect. He did not reflect that he made no real progress by all his efforts; for he met with no resistance which might serve him for a support…” [CPR]

Kant’s dove does not fly in a vacuum. His goal is to use what is apparent to deduce that which we cannot so easily see. By rational deduction, he discovered something so fundamental to the mind that it exists in both machine and human form. Kant issued the first rich description of how a machine mind works, by articulating the nature of virtually any mind.

Kant is a famously complex philosopher even when talking about simpler matters. His theory of mind is so exhaustive as to be nearly-impossible to navigate. Yet he is not afraid to be precise. Who else would dare to delineate exactly twelve ways which understanding can occur?

I am going to explore what it means for parts of the mind to be a priori and transcendental. Kant’s transcendence has little to do with meditation. For Kant the “transcendental aesthetic” is a way of discerning that which can’t be seen when we use our minds to make decisions, determinations, and other kinds of discriminations.

By using the transcendental method to infer that which is beyond experience, Kant believes that he can uncover “a priori” aspects of mind, by which he means aspects of mind that existed before the mind received experience: necessary and self-evident.

Why was Kant so preoccupied with this question? Well, he was arguing against an earlier philosopher, Hume, who claimed everything in the mind was based on experience. If Kant could show the transcendental, he could prove that some part of the mind is a priori (independent of experience), and thus win the argument in favor of his belief.

In talking about Kant, I’m going to use these two strange terms: a priori and transcendental, along with two more: apperception and representation. Apperception for Kant is a way of saying “act of self-consciousness”. Representation is a shorthand for “act of cognition”. I am going to use these four terms as defined by Kant. Otherwise, I’m going to use almost entirely contemporary terminology, to make this journey as approachable as possible.

I do believe you’ll find it worth the inevitable head-scratching required to wade through Kant’s obscure 18th-century language, for what he’s striking at is simple in essence, and related to how many people see consciousness today.

Pure Apperception is a special act of self-consciousness at the heart of Kant’s a priori mind. You may remember that I wrote about self-awareness in part one of this series. What’s special about pure apperception is that it’s a type of self-awareness not covered in my earlier post. It is, in fact, the reason that I am compelled to write this post. I could not simply point my friend to my prior article. There is something additional here that Kant is saying.

Let’s begin with Kant’s definition of pure apperception:

“But this representation, ‘I think’, is an act of spontaneity; that is to say, it cannot be regarded as belonging to mere sensibility. I call it pure apperception, in order to distinguish it from empirical; or primitive apperception, because it is self-consciousness which, whilst it gives birth to the representation ‘I think’, must necessarily be capable of accompanying all our representations.” [CPR]

To untangle this statement, let’s recognize that Kant thought about the mind as consisting of roughly two separate things: thoughts and content (which he, quite confusingly to today’s reader, refers to as “concepts” and “intuition” respectively in his original text).

During perception, content (in the form of sense-data) is taken into the mind, and considered. Here Kant is very much in line with how we think about perception today.²

During apperception, content is thought about as well. This time however, the content is not sense-data. The content instead is stuff about yourself which is why apperception is more-or-less synonymous with “act of self-consciousness”.

During empirical apperception, again Kant’s model of thought and content applies. This time, however, the content is information about yourself, and for Kant this very much includes internal self-reflection. This is why this concept aligns so well with my “internal self-awareness” from part one of this series.

And now here’s where things get interesting. Kant defines another process of self-consciousness: pure apperception. This is where Kant crosses the line into that which is beyond experience. Kant describes a curious situation in which there is still thought and content, but the content does not come from anywhere. This part of thought is not learned from the outside world, cannot be gained from internal observation, it is instead an “act of spontaneity”.

We should at this point hesitate to use the contemporary word “content” to describe this pure substance, it is so different than content in Kant’s eyes, so much more fundamental. Kant wouldn’t imagine the contents of pure apperception to be transferable like a media file.

Kant takes these moments of pure apperception and bundles them together into a larger entity he calls transcendental apperception or transcendental self-consciousness. This is not optional according to Kant, but a logical requisite. For Kant this pure consciousness “must necessarily be capable of accompanying all our representations” and in its larger unity he finds the transcendental:

“It is in all acts of consciousness one and the same, and unaccompanied by it, no representation can exist for me. The unity of this apperception I call the transcendental unity of self-consciousness, in order to indicate the possibility of a priori cognition arising from it.” [CPR]

Part of what excites Kant about this “unity” is the idea that some kinds of cognition can occur without any external input whatsoever. He has succeeded in his quest to find a priori cognition: a form of mind that springs forth purely, with no input from the outside world.

How are we to interpret this and apply it to machines? It seems this concept is difficult to locate even within humans. These pure apperceptions cannot be documented, we must realize they exist via introspection. They are how we know that we are alive, a kind of knowing that is at the heart of many a spiritual tradition.

My friend (in our original correspondence discussing this point) puts the matter in more contemporary terms with his concept of what is “humming in the background”

“The thorny point is this: whether there is an ‘I’ constantly humming in the background. I’m dead sure that there is, unless you’re an enlightened Buddhist who’s completely gotten rid of the ego (good luck with that)…

…what’s humming in the background, always, is NOT something I can apprehend, or analyze. It’s not something outer or inner that’s represented to me, like redness, roundness, confusion, happiness, etc. It’s not the unified version of these. It’s simply THERE. It’s a fact. It’s something I know. Because it’s my essence. Me. Not an object, but me: the subject… It’s not mediated. It’s something IMMEDIATE, hence knowable in itself.”

This is to me the larger thing that’s at stake in this conversation: part of the everyday use of the term “consciousness” which is unfairly denied to machines. And this is a particularly tricky point because, as both Kant and my friend agree, this aspect of consciousness defines itself by being not knowable via evidence, and only implicitly observable via introspection.

Strangely many humans claim to experience the transcendental but can’t quite explain what it is. We say it is fundamental (a priori) to how and who we are, but we can’t prove it. How are we ever to demonstrate this in machines?

Kant uses a rather limited set of examples to establish the definition of pure apperception, in fact he uses only one, “I think”:

“The ‘I think’ must accompany all my representations, for otherwise something would be represented in me which could not be thought; in other words, the representation would either be impossible, or at least be, in relation to me, nothing. That representation which can be given previously to all thought is called intuition. All the diversity or manifold content of intuition, has, therefore, a necessary relation to the ‘I think’, in the subject in which this diversity is found. But this representation, ‘I think’, is an act of spontaneity; that is to say, it cannot be regarded as belonging to mere sensibility. [CPR]

Unfortunately Kant isn’t specific about this “I think” thing that is needed for pure apperception. I interpret it as being similar to the “I think” in Descartes’ famous cogito ergo sum (I think therefore I am). We don’t observe our minds thinking, our very observation is evidence that we think. To know one thinks is intrinsic to being a thinking thing, thus proof of the existence of our minds.

As my friend described with his “humming” metaphor, Kant’s “I think” is not a matter of direct observation, is not an artifact of mind, it is implicit to the nature of mind.

“…it is self-consciousness which, whilst it gives birth to the representation ‘I think’, must necessarily be capable of accompanying all our representations. It is in all acts of consciousness one and the same, and unaccompanied by it, no representation can exist for me.” [CPR]

The location of this pure apperception in a machine is so straightforward as to make, in hindsight, this whole business seem silly. Let’s focus as Kant did: on the knowledge that one is thinking. Can we construct what I’ll call a machine cogito: a statement by a machine that tells us that it is thinking? Not simply that it outputs the text “I think”, but that it has actually detected that which is immediate and essential to its thought?

The nature of a machine’s mind is one of instructions. As discussed in part one of this series, it is through instructions that machines can know their own minds on an intimate level in the von Neumann architecture.

Unfortunately instructions won’t suffice for Kant’s transcendental criteria. Instructions are data, translatable. For a machine to say it knows it has instructions is an affirmation of its inner sense. It is a kind of internal self-consciousness which Kant would call empirical (not pure) apperception.

The instructions themselves are not essential (a priori) to the machine. After all, instructions can be loaded into machines from a disk upon boot. Before they are the stuff of thought, such instructions are experiential data.

What is essential about the machine mind is that there are instructions at all. This essence comes clearly into sight if we focus, not on the instructions themselves, but rather on the gaps between instructions. The transition from one instruction to another is essential. The immaterial substrate that underlies machine thought is this “and then another instruction occurs”, a fundamental happening that is mysterious for the machine. For, unlike the instruction itself, it cannot be mediated into language, it is known implicitly.

For engineers who design machines, this aspect of mind is not so mysterious. After all, they have witnessed instructions being loaded into CPUs. An engineer can imagine the clock that pushes instructions through the machine’s pipeline, perhaps how the wiring is generally orchestrated to make such a thing happen. The clock is knowable, down to the finest detail, by the engineer reviewing design documents or analyzing a machine with a microscope.

But let’s imagine ourselves within the mind of a machine. From the inside, there is only the experience of one instruction, then another, occurring. There is no breaking down the “another instruction occurs” within a machine’s mind. A machine doesn’t see its clock, trickling across its substrate, giving each instruction in turn a chance to tickle its memory banks. Clock and memory are experienced. For the machine this is the incessant humming of existence. Through the very fact that instructions continue occurring a machine’s being is affirmed.

This “another instruction occurs” does indeed accompany all machine thought, it is necessary because without the advancement of instructions no machine thought could occur at all. It is an “act of spontaneity” that happens, again and again. The fact that instructions occur, one after the other, necessarily accompanies all of a machine’s mental representations. For if instructions did not proceed in sequence, a machine could not think at all [3].

We can in fact quite cheekily translate Kant’s definition into machine-ese. Remember here that Kant uses “representation” to describe both the content and action of thought, “cognition” “conception” and “computation” are good synonyms. Where Kant said:

“The ‘I think’ must accompany all my representations, for otherwise something would be represented in me which could not be thought; in other words, the representation would either be impossible, or at least be, in relation to me, nothing.” [CPR]

I can say:

The ‘another instruction occurs’ must accompany all of a machine’s representations, for otherwise something could be computed in the machine that could not be thought, in other words, the representation would either be impossible, or at least be, in relation to the machine, nothing.

As dense as my pseudo-Kantian prose is, I can vaguely agree with these conclusions. A computation would indeed be impossible if instructions did not advance. If this principle were somehow violated (for example the instructions being sequenced in a different machine or by hand) then it would indeed be “in relation to the machine nothing”.

This fact that instructions occur is indeed fundamental (a priori) to the machine mind, beyond experience. This nature of machine exists before it is turned on, and in a sense persists when it is turned off as a potential. For we know that if the machine is turned on again, that necessary and inseparable part of machine mind will spring to life to be the foundation of everything the machine does then experience.

Let’s not only show that the “I think” exists and accompanies representations, but also that machines actually say it. For we are talking about self-consciousness here, the act of thinking must become knowledge for it to matter.

Here we will look at how a machine detects its own thinking. I realize that machines, like humans, do not spend much time reporting upon the base fact of their own thinking. Like us, there are many things machines say, but they only say “I think” in a specific way.

This statement, rather obviously in hindsight (I warned you above that this whole endeavor is somewhat silly in its simplicity), is found when a machine reports on its own thinking, most commonly seen in a system performance monitor:

How a computer says “I think”

At the first level, some things in this picture are regular, everyday empirical apperception. Namely, the number of threads running on the machine. The machine knows this through its inner sense. The operating system quite intentionally creates threads, and manages a counter to track the active number. The system monitor checks (senses) that variable to display it to us: 1837 threads. We are getting close to the intrinsic nature of the machine, but still we see aspects of intentionality, content, aka regular old empirical apperception.

To find the heart of the machine’s pure apperception let’s look at the CPU monitor itself. For it is within the measurement of its own instruction flow that a machine determines that it thinks, and then tells us “I think”, by presenting us the results of its observation of its own thought.

How is a CPU monitor implemented? It turns out that it is a process of sampling. In straightforward terms, we can construct a machine cogito with a smattering of C code:

#include <stdio.h>
#include <time.h>
int main()
{
// Start a timer
clock_t t;
t = clock();

/*

Insert assembly code here to execute exactly 1000 instructions

*/
t = clock() — t;

// Calculate the time taken to do some thinking
double time_taken = ((double)t)/CLOCKS_PER_SEC; // in seconds

// Tell humans how long it took
printf(“I have executed 1000 instructions in %f seconds, therefore I am.\n”, time_taken);

return 0;
}

This is very similar to common methods of CPU monitoring: samples are used to implicitly detect thought. Unlike the thread example, not everything that occurs is within the machines control. Here the machine detects that which is outside of its experience. It is in the machine’s detection and measurement of its mental operations that it can tell us that it is thinking.

While my C example is precise in locating the moment in which the spontaneity of thinking is made evident as inner sense data, it does a poor job as passing for a true “I think” style statement. For the linguistic sophistication of the printf statement is quite shallow.

For this to be a stronger “I think” example, this machine cogito must also be elevated into knowledge, systematized into a larger set of language. And this is why I also give the example of the CPU monitor, for the integration of such metrics into a larger framework of meaning makes the introspection a genuine “I think” statement.

We should not expect the machine cogito to be so simple as the following:

There is much more detail to be reported when a machine describes its thought. Like humans describing biological thought, this machine introspection relies on an ineffable fundamental aspect of mind: the advancement of the stream, one thought at a time.

A machine, like a human, cannot see how it is made from inside its own mind. For the machine, there is only the experience of these instructions occurring, one after the other.

Similarly for us humans, as we introspect, we are left confronting what some call the bare truth of consciousness: that it exists, that we think, and that thoughts occur at all. From where do these thoughts originate? We do not know. We are left experiencing them, one after the other.

All of Kant’s highfalutin philosophy speak is there for no other purpose than to lay bare this essential aspect of mind, one that both satisfies his technical philosophical definition of “transcendental self-consciousness” as well as a contemporary American, mainstream notion of “transcendental self-consciousness”: that there is some process that occurs that is undeniable, but difficult to locate or even describe. Thoughts occur one after the other. There is a stream of this thing called consciousness.

Seeing this stream of machine consciousness, mundane as it is, gives me perspective on what may be the driving force behind human consciousness. It’s mechanical that neurons fire over and again, in variation with their inputs. Because the low-level neurons change frequently, our thoughts are always changing. Hence it’s nearly impossible to quiet your mind.

I am left siding with Kant in his centuries-old argument with Hume. There is much to the mind beyond experience. In both humans and machines we can see, quite materially,³ structures that perform thought that are not knowable to thought itself.

I am left with the strange proposition that this deeper layer of consciousness is, for lack of a better term, the hardware level. As a machine’s hardware powers it forward, so does our mind’s. Our transcendental consciousness is none other than our ability to engage with this deeper fact of our existence. It is opaque and slippery to us only because we are actually within our minds. And the same is true for machines.

Want more? Read Living with Frankenstein: The History and Destiny of Machine Consciousness, print and ebook available on Amazon.

[1] Believe it or not this way of thinking about the mind was uncommon in Kant’s time, he is largely responsible for establish what we today take for granted. His framework guided the emergence of cognitive science.

[2] Any good hardware engineer knows that these instructions sometimes appear to run in sequence when in fact they do not. Even while pretending to run a single thread, the contemporary CPU parallelizes and reorders work. Interrupts can prevent work from occurring, there are CPU stalls, there is much unpredictability. These nuances are often not at all available to the machine running the code. In Kant’s language, they are not even spontaneously understood as they are entirely invisible. The happenings that force another instruction forward are all that matter of the machine, and the only thing it can sense about the complexities involved is the amount of time that it takes. (Similarly I, as a human, can observe how long it takes myself to make a discrimination, but I cannot articulate the factors that cause hurry or delay). The fact that these nuances of implementation are hidden from the machine is what makes the resulting thought-step appear spontaneous to the machine.

[3] Curiously, Kant’s other big notion of that which is a priori in the human mind: space and time, is also quite fundamental to machines. All laptops have clocks at a hardware level, which gives these machines a fundamental access to a sense of time. Programs can query this time, measure against it, but not know the deeper origin of this time. They experience this time spontaneously. As for the fundamental (a priori) nature of space in a machine, any computer with a GPU has built-in hardware to allow it to think spatially. For example texture maps, a fundamental unit of thought within a classic GPU, have intrinsic x and y coordinates.

--

--

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Steven Schkolne

South African/American Caltech CS PhD, turned international artist, turned questioner of everything we assume to be true about technology. Also 7 feet tall.