Artificial Intelligence has become too good. I don’t mean that AIs beat the best gamers or escape the sandbox during test runs. Rather, there is the remarkable phenomenon of Large Language Models exceeding the expectations of theorists and their programmers. No-one knows quite why the technology delivers so well. Statistical prediction of the likely next word in a sentence doesn’t seem to be explanation enough.
The surprising success is often taken as a cause for existential worry. “If anyone builds it, everyone dies.” But there is an alternative possibility.
The brilliance of language-based AIs may become an unexpected reminder of the mystery that pulses through you, me and - in one way or another - all things. Who knows? The advent of generative AI might prompt an awakening.
I was alerted to the possibility, in part, by conversations with a friend. Marius Dorobantu is Assistant Professor of Theology and Artificial Intelligence at the University of Amsterdam. He made an observation that is at once simple and disarming.
Recall that the big AI breakthrough came when the creators of LLMs stopped trying to parse language and feed computers semantic rules. Instead, they built data centres comprised of billions of words, including the context of their use, and left the LLMs to work things through themselves. ChatGPT leapt into the world and, alongside the iterations of the other big hitters since, an advanced AI future suddenly presented itself, promising both good and ill.
Mainstream commentators transmit the fear and hope in equal measure, a split that, to the psychotherapist, immediately suggests something in our understanding is missing. And this was the deeper point Marius suggested. What is overlooked is the significance not of the computational progress, but of the centrality of language to that progress and, therefore, to the importance of considering the nature of language itself.
Put it like this. The techbros and others in the field are shaped by a rational philosophy that underestimates the power of language, often out of homage to logic and maths. They are wedded to what philosophers call a correspondence theory of truth. This boils down to treating language as coded information that, at root, equates linguistic statements with facts about the world. “The cat sat on the mat.”
With this approach, the unreasonable success of LLMs is a puzzle. AIs can’t verify their outputs by checks with the real world, though the extraordinary quality of those outputs seems impossible without such testing. Hence, the rush to wonder whether the machines are waking up or whether a new kind of intelligence is being born. But there is a better, less melodramatic explanation that draws on a rival theory of truth.
The alternative relies not on correspondence but coherence. Consider what that means when it comes to language. For instance, to read, “The cat sat on the mat”, merely as an empirical observation that corresponds with a known state of affairs misses the point. The phrase is a play on the words that coheres across several dimensions. It describes something physically possible, yes, but it also coheres because of the rhythm (a bit of aesthetics), because it is fun to say (a bit of physiology), because it conjures a pleasing image (a bit of psychology), because it demonstrates a basic linguistic structure (a bit of epistemology), and so on.
In general, language works because it can embrace effectively an infinite number of multifaceted links, modes and relationships. The secret of LLMs, therefore, is that they are coherence engines - though unlike human users of language, who experience language, LLMs generate conversations and observations at one step removed from the experiencing. That is also crucial.
The outputs are more often than not true but also always artificial. Similarly, when it is said that LLMs might hallucinate that is also misleading. A correspondence model of language is being deployed with this explanation of mistakes - a “hallucinated” mistake being a correspondence with a world that doesn’t exist. But interpreted by the coherence model, mistakes actually mark the limits of relying on coherence alone, because uncoupled from experiencing language. Hence, whilst sophisticated conscious users of language run tests, such as whether an explanation has a ring of truth or whether a story passes the sniff test, LLMs are permanently stumped. They exist on the other side of a metaphysical abyss.
Properly understanding what is happening with the undoubtedly remarkable advances in LLMs, therefore, requires an understanding of language that says it is way more than nothing but a mode of information transmission. To move beyond that nothing-buttery, the intriguing connection made in the title of Marius Dorobantu’s job helps: Associate Professor of Theology and Artificial Intelligence. As Marius put it: if scientists are inclined to adopt minimal theories of language, theologians reach for the maximal.
In the Christian tradition, that means turning to the Logos - the divine Word running through all things as it creates, generates and communicates. (And, as if stressing the insight, many other wisdom traditions ensure a similarly central place for words.)
Language, on this account, in its fullest divine manifestation, is not merely a tool but a living manifestation of dynamic reality. When we speak, we share in that imaginative excess, which is why language can move us as well as be useful; it can open onto other worlds, as well as serve us practically in this one. When people love language, they participate willingly in an extravagantly rich reality.
Unpacking that, Rowan Williams has highlighted how the edge of words matters quite as much as the words themselves; he points to the intuitive, implicit and ineffable presence that accompanies words. Words are more than themselves - which is why a phrase like “I love you” is never overused but evergreen when uttered by someone who does indeed love you.
Alternatively, Owen Barfield argued that words have souls, which is why poetry works. Placing words in an unexpected order or conjunction releases that vitality afresh. “Shall I compare thee to a summer’s day?”
So, to return to the question of the unreasonable effectiveness of words in machine learning: whilst LLMs don’t experience language, they are astonishingly capable manipulators of the linguistic traces of that experiencing. That manipulation delivers more than might be expected because LLMs, in effect, piggyback on the power of language. Hence, too, an LLM can present fake feelings, seemingly follow purposes, apparently make imaginative leaps, but they aren’t doing so because they are conscious or intelligent. They are drawing on the inherent features of language, preserved in words, which catalyses the unexpected outputs of the data centre. The bots are so good because language itself holds all manner of relationships and links. Or, as a theologian might say: the echoes of the divine Logos reverberate even in the refrigerated silicon firings of a data centre. The algorithms don’t hear the resonances, but can amplify them - producing unanticipated effects.
There is an expression that any sufficiently advanced technology is indistinguishable from magic, but that is only because technological advances exploit and redeploy a magic which is already present.
LLMs can use language and do better than might be thought by virtue of the layers that are inherent in language. But still: they don’t speak it. So, whilst I suspect that there is no need to worry about LLMs becoming conscious, though there is of course a need to consider the ethics of their use, there are other concerns highlighted by the theological analysis which their deployment prompts.
For instance, might living with too much machine generated language be a bit like living too long under artificial light? There would be the LLM equivalents of mood changes, weight gain, sleep loss.
Similarly, an activity that depended too much on LLMs would weary: academic work rendered vacuous, fictional writing flat.
A tightly closed loop of artificial language production uncouples the language from the living source and so deracinates it. Or, to put it the other way round: the vibrancy of language depends upon its links to experiencing language, so if you uncouple language from that experiencing, it will tend to deaden.
Such effects are, of course, already widespread, a decline felt long before the advent of LLMs, when phrases or words that originate in abstract or technical contexts become common. Think of the aversion to managerial clichés because they dehumanise and also to ideological jargon because it substitutes for real thought. LLMs can extend, perhaps accelerate, that dulling. Remember when people started “delving” into things, deploying one of ChatGPT’s most distinctive verbs? “Delving” soon started to feel grey and people now groan when “delving” into this or that is suggested.
Highlighting these experiences is important. They awaken us to the mysterious spirit felt in language - negatively when the language has become empty, positively when stirring the longing for it once more to live. There may even be a stirring to ponder anew a third thing: the ever-present origin of languages - the Logos. The artificial might help the rediscovery of the real, alongside the wonder of being language users that modern worldviews tend to sideline, meaning people forget.
To put it maximally: human beings participate in the image of God, an image found in all created things and capable of being distinctively known in the use of language.
An LLM is the Word made digital abstraction. But that can prompt a desire to know the Word made flesh.
DATES FOR YOUR DIARY
Saturday 5th September - Spiritual Fourfold London: A William Blake day pilgrimage with Mark Vernon. The first of several events looking forward to 2027, the bicentenary of Blake’s death.
Thursday 17th September - Awake O Sleeper in the Land of Shadows: William Blake’s call to us now. An evening with the Temenos Academy looking forward to Blake 200, too.
Thursday 8th October - Idler Drinks with Michael Pollan. The first of the new season of online weekly gatherings.
Saturday 17th October - Spiritual Fourfold London: A second William Blake day pilgrimage with Mark Vernon. Another chance this year to prepare for the bicentenary of Blake’s death.
Friday 27th November - Escaping the Cave: Plato in Art, Science and Spirituality. A day conference in Cambridge.
Thanks for this very well written, lucid, and hopeful essay. Brings these words to my mind: "The letter kills but the spirit gives life."
Killer ending!