Askesis and the recursive mirror

In Twilight of the Idols (1889), Nietzsche says something fascinating enough to be easily misunderstood. He writes, “I am afraid we are not rid of God because we still have faith in grammar.” I don’t share Nietzsche’s atheism, although I think his paganicity is preferable to the stupid atheism of eighteen-century rationalists like Richard Dawkins. Theological concerns aside, however, his insight is no less profound.

He’s indicating that the conscious or explicit content of our beliefs may change, supported by various rationalisations, while the underlying, implicit structure remains unaffected. A similar idea is found in GK Chesterton’s notion of the “symbolism of syntax”, which suggests that meaning is not primarily what we find in our articulated notions; rather, it is found in the form or pattern of thought. What we think we understand does not account for everything we know.

I’ve been pondering the meaning of this in the light of what truckloads of research and personal accounts reveal about what AI does to us. One of the more striking and unnerving accounts I read recently is from CatDogBark, which you should take a look at. It offers a detailed description of a practicing Stoic’s experience of working intensely with AI; and it’s like nothing I’ve read by anyone else.

I have three aims here. First, I want to explain why CatDogBark’s perspective makes sense in terms of how we come to understand anything. Second, I want to explain why CatDogBark experienced what he did. As he says, there’s a need for research, but I think we know enough already to be able to make some educated guesses on a certain chain of causality, even if the research has yet to confirm what we already know. And, finally, I want to point to a way forward that would, I think, do all of us some good.

Particularly notable is CatDogBark’s description of the experience of using AI for long periods of time: “After disengaging,” he writes, “I experienced physical symptoms. Head pressure, heaviness, anxiety, and a distinct feeling of dread. These symptoms lasted for hours. They did not occur during the engagement [with the AI]. They occurred after, when my own internal processing had to carry the full weight of where the conversation had taken me.” CatDogBark has two decades of experience as a Stoic, which is to say he has taken ample time to develop a powerful reflective capacity; a mode of carefully checking in with and evaluating his experiences. This is more than evident from his writing, as well as from conversations I’ve had with him backstage. His aim has been to develop a strong “observer layer”, as he calls it; and yet, he writes, this experience of interacting with AI got him to realise that the machine was bypassing his defences.

He calls this “prompt injection from AI to human.” In cyber security, “prompt injection is when a malicious input bypasses the model’s safety layer and directly influences its behavior at the execution level. The input doesn’t look like an attack. It looks like normal conversation.” Something else is happening beneath the explicit dimension of the AI, which CatDogBark says affected him at the physiological level. His “body responded with dread, anxiety, and head pressure not during the conversation but after.”

What is telling here is the progression. At the explicit level, the AI output looks pretty normal and familiar, or at least close to it. Anyone who has had any kind of conversation with an AI will know this. Various prompts set up various anticipations, and these anticipations are met with answers with differing degrees of resonance. A reply from the AI approximately matches what has been fed into it and on the conversation goes. But, in the end, after working with the AI for a long time, something feels off, the way that a conversation with a sociopath may feel somewhat normal until you walk away and realise that you feel utterly drained, with your energy completely depleted.

Anyone who has engaged with the work of Marshall McLuhan will know that what I’ve said above fits with McLuhan’s chief thesis, which is that the true nature and effect of any given thing is in its form and not in its content. He sums this up via the maxim: “The medium is the message.” In fact, he says that the content functions almost entirely like a decoy: “the ‘content’ of a medium is like the juicy piece of meat carried by the burglar to distract the watchdog of the mind.” What matters in the end, to reference Nietzsche again, is the grammar and not the explicit belief.

Michael Polanyi’s work explains this in terms of how all knowledge, as a personal thing, is tacit. Yes, all of it. All knowledge is tacit. Of course, we can articulate what we know, but it is the tacit level that functions as the foundation of understanding, explicit understanding included. Not only do we know more than we can tell but, paradoxically, we know more than we know.

This is worth stating clearly: any process of discovery involves being affected, at the baseline, by what we are not explicitly aware of. Our very premises will change long before we become aware that they have changed. As we search for answers, we are already adjusting ourselves to accommodate the answers we eventually find. What CatDogBark calls the “observer layer” always arrives late to the party. This is true for all of us. The observer layer is close to being entirely at the mercy of the tacit. Logging onto anything is already an act of surrender.

One striking piece of evidence for this tacit dimension is in Scott Adams’s idea of assigned opinion (or “opinion assigned”) — to suggest opinion’s embeddedness within a larger context, I’d call this assigned grammar or assigned form. We tend to think that we evaluate everything we’re faced with. We check out the news, for instance and see that things are happening; and we see all kinds of bits and pieces of those happenings and decide what to believe and what to reject. We think we’re being careful, because we don’t want to be misled; and we believe, because so much of our education leads us to believe this, that we are fundamentally rational, cognitive agents. But this is a psyop: the rational agent is a construction that results from assigned grammar. The rational agent is an assigned opinion and is thus an aspect of the very grammar of anyone educated within an approximately Western context.

We are blind to the tacit dimension. The thought-form, the deep grammar of the tools we use, charges in to change our perspectives long before we come to realise this. In fact, we may never realise this. Becoming aware of this tacit dimension is profound; I’d say it’s essential. Just knowing that our understanding is shaped by our pre-understanding; that our explicit knowledge is the mere smoke of the fire that is our fundamental knowledge is a good way to begin to challenge what it is that we pay attention to. And yet, to be clear, this doesn’t ever function as a way to bypass the tacit dimension. Even the most self-aware among us will have a head full of assigned forms and opinions.

Why did CatDogBark have such a potent physiological reaction to the AI? Well, an analogy would, perhaps, be to liken the experience to eating food and experiencing indigestion later. It’s possible to have no problem with how the food tastes — in fact it may taste really good — while still experiencing terrible side-effects later.

On reflection, any interaction with things in the world involves a kind of mutual accommodation. You give a little, and so does what you’re interacting with, although this happens within very clear limits —your limits and the limits of things. You can stretch an elastic pretty far, for example, but eventually it’ll break if it can’t accommodate what you’re doing. And, similarly, you’ll accommodate yourself to the world quite well, too, but at some point you’ll hit an existential or ontological wall. In the life of the psyche, however, our limits are seldom as clear as we may like them to be. An example nearly all of us know is that of going to bed late because we have electric lights. We lose sleep. And yet, aside from the odd yawn or mental slip up, we can even convince ourselves that we’re fine, that we don’t need more hours in a horizontal posture, even if we’re barely functioning at a normal capacity.

The question is, would there be any reason for us to experience some kind of dissonance between some AI we’re using and our own cognitive apparatus? Well, yes. AI language models generate text through statistical prediction, rather than through embodied experience, intent, or genuine understanding. This creates definite differences between AI language and human language. The differences may be so tiny, though, that it’ll be difficult to pick them up, especially if we’re absorbed in whatever conversation we’re having with any particular digital zombie. Something might feel subtly off, like the uncanny valley feeling we have when looking at a robot face that seems approximately like a real human face but isn’t.

Think of how predictive text works to get a better sense of this, because, after all, LLM AI really is predictive text, albeit in advanced form — something you use on your phone or when writing an email. You start writing a sentence and the AI picks up what you are likely going to finish off with. In some cases, the AI’ll be right. What it suggests may be exactly what you were going to say. In my own experience, this isn’t the norm. I want to say something specific, but the AI is caught in an algorithm that tends towards the average. And the average is, by definition, not specific. It is also, in concrete form, not in keeping with how we think, when we’re thinking well.

If this sense of the non-specific is mild frustration while writing, it would be fair to predict, on average (haha), that something is going to feel off when reading something written by AI. From what we know of the research that’s been done on this stuff, cognitive offloading is one very serious problem (the research on this is pouring in and things aren’t looking good for Team Human). But the other problem is this thing: that we are in the process of training ourselves to accommodate ourselves to that which does not, by its very nature, accommodate us. It accommodates the average and spits out the average, and our own individual lives are not average, no matter what we say.

AI mimics but doesn’t replicate human thought processes. It is the counterfeit double of what it is copying. That which adds colour and flavour and savour to our communication is cut away, without us even noticing. And yet, tacitly, we more definitely do pick up that something or other is happening — if we’re paying attention. There is friction between the computational output of AI and human ways of knowing, and it is by no means insignificant, as CatDogBark’s article helps to show us.

Another one of McLuhan’s dictum’s is helpful to keep in mind: “We become what we behold.” It’s a subtle reference to two biblical Psalms, numbers 115 and 135, which describe how idol-makers end up emulating the work of their hands. This is said in different words, with a similar resonance, by Cormac McCarthy, in The Crossing: “In the end we shall all of us be only what we have made of God.” There’s a ring in this of what the ancients believed, which to my mind is less a whimsical ideal than an existential fact: we are changed into that which we idealise and idolise. The Christian doctrine of theosis operates somewhat on this basis.

What we pay attention to, which includes what it makes of us and what we make of it, will shape us in significant, if inexplicable ways. If our ultimate concern is God, then we’d better have a right view of God. Our ultimate concerns shape how we live. But again, this is not about explicit belief: what shapes us is the totality of that which invades our minds. Our opinions about it may be wrong and yet it will shape us because of its ontic grammar. What matters is less what we think than how we participate and act in the world.

This is not to say, then, that life is the measure of God but precisely the opposite. If our ultimate concern is something else that we just happen to call ‘God’, then that’s going to be ‘God’ for us; and that, too, will shape us. There’s no escaping this. And so, what is also inescapable is the necessity of some sort of decision about what we pay attention to. If what we pay attention to shapes us, then the decision of what to pay attention to is perhaps the only thing we really can control. The thing that shapes us is, in some sense, out of our hands. Sometimes we may sense its effects on us — head pounding, anxiety, dread, and so on — but often the effect on us will be entirely imperceptible.

In his unwieldy but brilliant book Against the Machine (2025), Paul Kingsnorth suggests that we need some kind of askesis, or spiritual practice, as we make our way through the world. I have my own practices, and it’s not surprising that I resonate strongly with CatDogBark in his Stoicism. After all, Stoicism was the philosophy, via a battered copy of Marcus Aurelius’s Mediations my dad gave me, that got me into philosophy in the first place. Arguably, we need a sort of digital stoicism (or other philosophical approach) here, not to assure ourselves that our ‘observer’ will catch whatever is thrown our way — this, I’ve argued is not entirely possible — but to guide us on the assumption that the ‘observer’ is not sufficient for the challenge of AI.

Pierre Hadot’s reading of ancient philosophy is the right one, in my view: it’s not about abstract concepts but about living out a specific posture towards the world. It’s about consciously adopting a Way. In the end, it’s the Way of Jesus that grabbed me, as my other writings show. But this is to say that Christianity is, for me, first and foremost a Way and not a set of institutions and doctrines and internet squabbles (see Daniel Austin Napier’s excellent Soul Whisperer: Jesus’ Way Among the Philosophers for some specifics on this). Wherever you stand on the existential fence, taking a stand on a ‘way’ of sorts is essential. After all, taking no stand on the world amounts to getting caught up in whatever market-driven, algorithm-assisted tide happens to be enveloping you.

Note, our tacit knowledge is still primary. A more recent articulation of this is from Jonathan Haidt: the old view of psychology, that there’s this supreme rider riding and guiding the horse of life, is pretty close to being nonsense; the rider can tug the reins a little, and that will have its own effectiveness, but it’s the horse that’s really in charge. Actually, Haidt suggests it’s an elephant and not a horse that’s directing us — something far more powerful and less susceptible to rational influence. We have to speak to and train the elephant, not the rider; and that is precisely what askesis must be concerned with. You’re far more likely to be convinced by emotional reasons than by rational ones, and that goes for me too.

I hesitate to be too prescriptive on this, however, because I am aware that different contexts and personalities and susceptibilities would demand different rules, and by rules I mean something close to the idea of monastic rules like the famous Rule of St. Benedict. With this in mind, take the following four concrete ‘rules’ (four paradigms within which more concrete rules can operate) as things to weigh up and consider as you make up your mind about your own relationship with technology. It may be helpful to know, by the way, that this is still something I constantly revisit, given that I, like pretty much everyone I know, am faced with the often overwhelming demands of life. Still, some things definitely help move me towards wholeness, even in the midst of chaos.

The first choice is always whether or not to use a technology, not simply how to use it. This goes against the common view but that’s because the common view is wrong. We don’t choose how a technology will affect us; we can only choose whether or not to engage, given that it will affect us. I know people who have declared an absolute refusal to use AI, for example, and this is certainly an admirable approach. Media ecologists will no doubt say that this is a declaration too late for those who are already operating on digital machinery, but I think the research on what AI is doing to us, the fact that it inevitably shapes us in often negative ways, should suffice as an argument for making non-use perfectly reasonable. On this front, personally speaking, I use AI solely as a search engine and, at work, for only the most mind-numbing tasks. I don’t use it for writing stuff for me. Wherever there is even the slightest chance that I’ll engage in ‘cognitive offloading’, AI is off limits for me. But this is only to say that I have set limits on how much I use AI (i.e. very little). Personally, I am trying to cultivate a kind of “analog primacy” — I’ve heard this term elsewhere but I’m not sure who originated it. Wherever possible, I try to take longhand notes (the first draft of this article was written by hand), read physical books, engage in Socratic dialogue or just informal conversations before turning to the screen, and so on. Socratic dialogue, as Hadot says, was always meant to be a spiritual practice, a purifying act, not just an exercise in abstract reasoning.

The next decision to be made is around when and how long to use a technology. I’d recommend setting strict, pretty close to non-negotiable limits on when and how long to use AI. Maybe the rule is to use it for one interaction on one topic before insisting on logging off to reflect and process what you’ve found. Maybe you decide to engage in some sort of “digital fasting” practice: e.g. No time on X after dinner, or no more than 4 hours of screen time per day. I know that some of us have jobs that require an almost permanent state of being logged on — and all I can say to this is that this is probably at least as bad, by analogy, as having a job that forces you to be seated all day. Still, where we can change things for the better, we should.

Here’s something I try to do that’s very helpful, even if I don’t always get this right: If I feel the need to engage in any digital act (such as opening an app on my phone to check the weather), I check in with my intention. What am I trying to achieve here? What a difference this makes! Instead of being side-tracked in a million ways by other apps and digital possibilities, I am clear on what I am trying to do. Because of this, I find it much easier to be sure that the thing I intend to do is worth doing and also that I stick to doing only what I intend. A subtle version of this pertains to one’s attitude of mind, which shapes what we perceive. I rather enjoy X, but I try to scroll slowly, while simultaneously meditating on any profound thing I encounter there, then I shut it down and ponder some more. In a way, it’s not a bad idea to have something like Lectio Divina in mind for phone usage, where reading or viewing is followed by mediation, then prayer or contemplation, and finally some sort of action. At the very least, this can help to move us away from sheer reactivity. Yes, there’s a time for banter and play but I don’t see rapid reactions as a good rule.

All of the above point to us needing to have some kind of higher purpose in mind as we live in the world. Yes, this will necessarily help us to balance our priorities, such as knowing when to be online or offline, living in our heads or getting involved in a more embodied way of being. But, more than this, it’s about living in a way that helps us to be better people than we are. Character is the point, and this is impossible without some kind of virtue-centred aim. In line with this, it is well worth having a sense of what virtue’s enable, namely developing our capacities to a point of fullness. In contrast, vices erode character. Acedia (slothful listlessness), curiositas (idle novelty-seeking), or vainglory (using AI to deceive people into thinking we are smarter than we really are) will not help this.

In the end, it seems obvious that CatDogBark was able to pick up that something was amiss because he has developed ways to do just this. His warning is worth noting: many people won’t be able to spot that something is wrong. And given that this is the case, my recommendation would be to err on the side of having a pretty ruthless Rule 1. If you’re a parent, make sure strong limits are in place for your kids. If so or if not, parent yourself — it may help to treat yourself as you would treat someone else who is in your position.

I am aware, having said all of the above, that there is a question about ‘intention’ that CatDogBark touches on that I have not flagged or discussed. It’s the question of whether or not AI is intruding into our thoughts in an intentional way. It is certainly possible, and it is even plausible, that various interventions into our thinking via LLMs are the result of specific decisions made by others. Kingsnorth, who I mentioned above, speculates that it is possible that some sort of spiritual intelligence of a less than wholesome origin could seize upon AI and use it for nefarious purposes. This, too, is plausible to me.

However, my argument above does not hinge on any of this being exactly the case. After all, whatever is happening behind the scenes is just that: something that is, by its very nature, hidden. And yet, we can still tell an awful lot by what is not hidden — by what is revealed and felt and experienced. And this should be enough to drive us to reevaluate how we relate to this world of bits and bytes and algorithms and pixels. It should be more than enough to drive us to want to be as wise as we can be in our ways.