I am pleased to share with readers of Tradition & Sanity the following incisive guest article by Dr. Jeremy Holmes, Associate Professor of Theology at Wyoming Catholic College. As I do from time to time with articles I consider especially important and deserving of the widest possible dissemination, I have published this post without a paywall. Enjoy! —PAK
The public debate over Artificial Intelligence rightly focuses on whether AI is, on the whole, a good thing or a bad thing. At one extreme, the lords of Silicon Valley assure us that AI will save us from all ills, free us from all work, and replace us as the dominant species on the planet. These take as their standard Genesis 6:4:
COME, LET US BUILD OURSELVES A CITY, AND A TOWER WITH ITS TOP IN THE HEAVENS!
At the other extreme, culture warriors warn us that AI will make us stupid, enslave us, and replace us as the dominant species on the planet. These take as their banner Joshua 8:28:
SO JOSHUA BURNED AI, AND MADE IT FOREVER A HEAP OF RUINS, AS IT IS TO THIS DAY.1
This essay joins the quieter conversation in the background, where people are asking: What is AI? Is it “intelligent” in the strict sense of that word?2 More specifically, I want to look at what most people think of when they hear talk of AI, namely Large Language Models: ChatGPT, Claude, Grok, and all the rest. AI taken broadly seems to fall under Aquinas’s oft-repeated dictum, “Every agent produces its own like”: in the realm of artifice, man’s production of his own like reaches a kind of zenith in AI, which appears to resemble man in his inmost secret, namely intelligence itself.3 But LLMs can be situated more precisely as a next step in the history of man’s relationship to language.
Just as we are embodied spirits, so we embody our thoughts in language. When our thoughts are purely intellectual, in no way enfleshed, they are elusive even to us, but language helps us get our thoughts out in front of us, so to speak, where we can examine them and play with them and reorganize them. Those unfortunate enough to grow up without language seem like animals, unable to tap into their own rationality. Language is built into our nature as the proto-tool of thought, as hands are the proto-tool of action.
However, readers of this essay encounter language under many layers of technology: writing, uniform printing, digitization, Internet. Walter Ong, S.J., writes in Orality and Literacy: The Technologizing of the Word that “Fully literate persons can only with great difficulty imagine what a primary oral culture is like, that is, a culture with no knowledge whatsoever of writing or even of the possibility of writing.”4 We have to place ourselves imaginatively in a situation where words have no visual aspect: they can be “recalled” or “called back,” but you cannot “look them up.” They are not things or objects and they leave no trace (a visual metaphor): they are simply events, happenings. They pass away even as they arrive. This is why the Hebrew term dabar means both “word” and “event”.
So in the purely oral situation, if you want to keep the benefit of words then you just have to remember them, and if you want to remember what you said then you have to say memorable things.5 A purely oral culture puts a lot of energy into saying things memorably: it possesses an enormous complex of mnemonic techniques and habits, and repetition in this context is a virtue. The purely oral speaker is immersed in his words like a fish swimming in water; language pushes thought out in front of him, but not in any stable way. His speech knows nothing of an abstract, neutral list or a how-to manual, because his thought stays close to the situation of human life. To be thinking is for the language event to be happening and happening in a beautifully memorable but still fluid way in moment-to-moment dependence on the speaker.
With the advent of writing, the word stands still.6 It gets literally “out in front of” us, and it stays put, having become of a kind of “thing” at last. As a result, memorization is not the only way to recall what one said before, and the whole complex of mnemonic techniques loses urgency. In a manuscript culture, thought processes are more linear, repetition less valued, and the experience of language becomes more detached. Someone who has grown up writing can hardly imagine a word without spelling it in his head: even spoken language has become more thing-like.7 Our ability to put our language in front of us and analyze and examine it soars, bringing with it the art of logic properly so called as well as theories of grammar. The act of language becomes more solitary, because while spoken language would rarely happen in the absence of anyone to hear, writing normatively happens without the audience actually present. So language becomes in a way more interior, more something that just happens “in my head”.
The move from manuscript to printing press doubles down on all these trends.8 The word is even more a “thing”: not only does the word stay still on the page, but all the letter shapes pre-existed this particular assemblage of letters, and every copy of a book is identical to the others, less tied to any one scribe and more just its own thing. Frodo may have met his friends at “the sign of the Prancing Pony,” but that was in manuscript times; in the printing press age the painted pony gives way to words, such as “Lander Bar”. And people read “Lander Bar” silently, without evening imagining a sound in their heads, as they fly down Lander’s Main Street at speeds unimaginable to hobbits, because the word has ceased to be an object for the ear and has become instead an object for the eye alone. Books are available cheaply, so there is less reason than ever to memorize; the printing press ushers in the golden age of the dictionary and eventually the encyclopedia, and a gentleman can sit in his private library with the knowledge of ages around him at his fingertips. Reading becomes normatively silent, an entirely solitary affair. Descartes retreats into his cabin, and the world goes mad over the possibilities of reason.
The digital age carries these trends even further. Computers brought to writing the same standardization that the printing press had brought to printing, so that even as I type this essay the letters all pre-exist in some way in the computer and I just call them down by pressing the appropriate buttons. There is no longer a “manuscript,” literally a “written by hand,” a personal physical link between author and first draft. Moreover, the final result is even more distant from the one who would read it. On this point, Byung-Chul Han’s work Non Things is worth quoting at some length:
An e-book is not a thing, but information; it has an altogether different status of being. Even if we have it at our disposal, it is not a possession. It is something to which we have access. An e-book reduces a book to informational value. … Information has neither a physiognomy nor a fate, and it does not allow for the formation of intense ties. One cannot have, for instance, a personal copy [Handexemplar] of an e-book. A personal copy of a book is given its unmistakeable face, its physiognomy, by the hand of the owner. E-books are faceless and without history.9
While the digital text’s quasi-immaterial status means that it cannot precisely be “possessed,” the same status means that unlimited numbers of readers can “access” it. The gentleman mentioned above no longer needs a dedicated library room, because the entire literature of all the world is accessible from his hand-held device. Incentive to memory vanishes entirely, and even the incentive to think softens because a quick search will pull up how someone else has solved any given problem.
At root, language works by giving thought some existence outside our spiritual power of knowing. Step by step, we have “technologized the word” by giving it more and more autonomy from us, seeking more and more benefit from language. It all started with getting thoughts outside of ourselves as events in the air, through speech. Then we got our thoughts out in front of us as stable things, through writing. Then our words became even more thinglike and independent, through the printing press. Then all of our words became pseudo-spiritual beings beyond the grasp of any hand but available as a kind of army at the touch of a finger, through digitization and the Internet.
Large Language Models take the next step. With LLMs, the texts we have written vanish into the inaccessible realm of “training data,” as we push not just words but language itself outside of us. It is as though we meet our whole faculty of language, there in front of us, speaking all by itself. We keep making the language act more and more independent of us, more a “thing” on its own, and eventually there it is: language itself steps out of us to turn back and greet us.10 As language has always been the handmaid of reason, it is no surprise that language externalized seems as though it were rational. Previous progress in technologizing the word made knowledge ever more available and so disincentivized memorization; LLMs respond to our questions directly, disincentivizing not just memory but the reasoning memory always supported. Seen this way, ChatGPT and Grok and company constitute a remarkably smooth continuation of what we have been doing over the past several thousand years.
At the same time, there is an equally remarkable rupture with the past. No matter how much we thingified our words, they were still our words, and they had to happen in our minds before they could pass into any more independent condition. LLMs sever the link between language and mind: for the first time, there are words but no speaker. If this seems the next logical step, it is also a leap like nothing before it.
Consequently, I have to circle back to question my own claim: Is that really language greeting us in an LLM? Thought is to word like soul to body. As the soul expresses itself tangibly through the body, thought expresses itself imaginably through words. How can LLMs possess language when they sever its connection to thought? But on the other hand, if LLMs do not possess language, what is this in them that seems so much like language? How can we describe an LLM so as to capture both the continuity and the rupture with past language technologies?
I find it helpful at this point to introduce the concept of a philosophical zombie. This is a purely theoretical fiction used by philosophers to develop arguments about consciousness, but the definition of a “philosophical zombie” is something that is physically identical to a human being but lacks any conscious experience. That is to say, every molecule and every atom and every electron in a “p-zombie” is exactly where it would be in a real person and is moving exactly as it would in a real person, but there is no intellect, no will, no awareness, no love, no hate—no soul. Supposing you met a p-zombie and spoke with it you would think it human. Supposing you met a p-zombie that you knew beforehand was a p-zombie, you would reap the stuff of nightmares: not a who but an it, not a he or she or even a “they” but just an it, like a Gollum out of Jewish folklore but with a precise resemblance to your neighbor, creating vocal sounds identical to words and touching you with that warm but ultimately oh-so-cold non-hand.
Can such an abomination exist? I don’t think so.
But with LLMs, it does seem that we have a philosophical zombie of language: it is physically identical to language, behaves like language, responds like language, and yet has no thought or consciousness behind it. Physical identity but no “soul”.
LLMs work by first “tokenizing” language and then transforming tokens into vectors. That is to say, they first identify the little chunks that language is made out of and then turn each little chunk into numbers, so that they can create a kind of map in which any given “token” or language chunk is near to or far away from any other given token. In other words, LLMs work by transforming language into relationships between quantities. While their success is startling, in retrospect it makes some sense, because all language traces back to spoken language, and Aristotle long ago described spoken language as a kind of quantity. Our thought enfleshes itself first of all in quantity. So a machine that recreates language as quantitative relationships has in fact come in at the ground floor, so to speak, tapping into the basis of everything else in speech.
I will not plunge on into the debate about whether LLMs are good or bad. However, it will surely be helpful for that debate to see that LLMs are the next step in a process that has been going on for millennia, so that we can look to the past to see how previous changes affected us. It turns out that writing, the printing press, and digitization each had profound effects not only on how society is organized but even on how we think—on what we think thinking is.
These past experiences can offer principles for thinking about the present. Moreover, while everyone senses that AI brings something new, it will be helpful to see more precisely what rupture LLMs bring with them. This, too, should help us think about what future the new technology offers or does not offer, and what role our society should assign to it.
The original scripture verse does indeed say: “Joshua burned Ai.” Ai was a place in central Canaan, first mentioned in the Bible in Gen 12:8. More on this town may be found here.
On AI taken broadly, I especially recommend the podcast series by Michael Augros titled “The Mind and the Machine.” https://www.thomasaquinas.edu/mind.
See, for example, Summa theologiae I, Q. 1, art. 3, obj. 2. But if you put “sibi simile” into the search bar at aquinas.cc you’ll see that he repeats the same principle verbatim again and again throughout his works.
Ong, Orality and Literacy: The Techologizing of the Word (Routledge, 1988), 31.
Ong, Orality, 34.
For this paragraph see Ong, Orality, chapter 4.
My children often pronounce words correctly as long as they have never seen them in writing, but then mispronounce them incorrigibly from the moment they see them spelled out on a page. That is to say, from that moment on they never speak the word without reference to a visual memory of that word.
For this paragraph see Ong, Orality, chapter 5.
Byung-Chul Han, Non-Things (Polity Press, 2022), 16.
To be clear, this is not “language” in the sense of a dictionary and a grammar book, but our language as it has actually existed, forming and being formed by each new text composed.
We seem to be nearing things like this:
“And it was given him to give life to the image of the beast, and that the image of the beast should speak; …”
(And for the sake of completeness):
“…and should cause, that whosoever will not adore the image of the beast, should be slain.”
Apocalypse 13:15
Douay-Rheims 1899 American Edition