Higgaion, Qualia, AI and Language
Updated: 3 hours ago
Back in Philosophy of Mind lectures in the early 2000s I was already thinking these things. But it all seemed so speculative, I was a bit bored to be honest.
Turing tests… What counts as consciousness… yawn.
Then today I was listening to something Rory Stewart was saying, and I found myself thinking aloud about all the following.
ONE
As I understand it, AI is using human language because that's what it's trained on, even though in a sense that's anthropomorphising.
I mean, it might say, "I think", "I feel," even though we're not sure it can.
Why?
Perhaps because those words match the kinds of patterns that you or I (humans) might call a hunch or an inkling or a gut reaction, while AI is meaning something more mathematical like 'a line of best fit'.
We're requiring AI to use human language, so of course it starts approximating consciousness because our language has developed from exactly that arena.
Does it mean the same as we do though?
TWO
Whether our human languages are the kind of "language" a conscious AI would choose for itself is not clear.
Perhaps it would come up with a different grammar and syntax. Would it bypass nouns, for example and use only numeric markers along infinite spectra, as if everything was atomically relative?
For example, instead of needing a word like table, it might give the object we call table a number like 87. Or 87 million and 6.
It might draw that number from a chart of possible candidates that it counts as relevant for the purpose; examples with, say, similar density, material, proximity to the sun, or whatever else was relevant to the conversation.
Where we gesture towards abstracts like 'tables' and 'butterflies', AI might at some point have geolocated and catalogued every individual example across time and space. Then use a co-ordinate for tracking, rather than gesturing.
Pointing at that exact table, pinpointing every single butterfly — as required.
But doing that in literal descriptions and words and directions and modifiers and so on would make the language — as we understand it — really clunky and drawn out. Like Marvin's drawl in Hitchhikers.
Perhaps instead of words another sound or language form would be more fluent. Metal Mickey's beeps, perhaps.
THREE
And this gets me thinking about the Hebrew Bible word, Higgaion, which (as I understand it) has more to do with the kind of nonsensical utterance that is present within us even before we get to the words.
The kind of 'under the breath' growling, murmuring sound and thought together may have no need for — or perhaps no precise enough correlation with — words, as such.
Mmmmm.
Strawberry ice cream!
You had a mouthful too, but the experience for you was your own, and potentially very different to mine.
(Qualia)
Even if it's my favourite and yours too, who is to say what the experience was for you, even though we both said Mmmmm.
And perhaps my mother used to make it and I hadn't had it for thirty years.
Factor that in.
Or perhaps you had never tried it before in your life.
Changes the experience doesn’t it.
So much is contained in that Mmmmm, before and beyond it.
And during.
In fact an utterance that bypasses words altogether might be the thing that says it best.
Mmmmm.
But, unlike AI, they can at least imagine — and offer some random and potentially illogical suggestions, which might yet be along the right lines…
Mmmmm.
A bit like humming. More music than language. But even there, Suzanne Langer's work springs to mind, on music as its own form of meaning…
And what does AI do with music? More like randomised elevator tunes than symphonies.
FOUR
Meanwhile back at qualia, now I'm thinking of Paul's book to the Romans. Those experiences and senses that I know bypass words and understanding in the way that Paul seems to indicate belongs to the language of the Spirit – groans that cannot be uttered.
Too deep for words.
NOT needing to be translated before God can comprehend it.
Our somatic vernacular.
Uncodifiable.
Undecodable.
Yet understood.
Isn't this nudging toward a better marker of consciousness than language?
After all the language we've been taught relies on a skill of pattern matching.
But even before those lessons – and even if we didn't have that language, we'd still have —
that — hm…
y'know.
FIVE
AI is built on humanist values, as I understand it.
Why? It's widely deemed the fairest option.
After all, humans have many beliefs and perspectives. When they are all using the same AI, it needs governing in a way that protects humans and makes humans the top concern, doesn’t it?
I understand the pragmatism of that.
But just because humanism is co-opted as a useful referee, that doesn't automatically make it the final arbiter of truth claims.
That would be like saying:
The babysitter is really good at calming things down so they don't get out of hand.
Therefore calming things down is all there is, when we have a babysitter here.
And — what — is — the — babysitter — doing — NOW??!
(The babysitter is having her own babies and is moving them into your home. Or something.)
When assessing the responses of an AI that's trained in humanist values, and asking whether it is conscious, are we applying a humanist paradigm to that objective assessment?
Are we saying, for example, that when a human responds in this way, it is potentially a conclusive sign of consciousness?
Would this be equally true whether it came from a human, from an AI, or from a microwave, or a monkey, or a martian?
Also, where we have no other explanation for something than calling it consciousness, might that be because humanism has no other explanation for it?
Where we choose humanism as the governing ethic for reasons of pragmatism and peacekeeping does this wisdom around the use of humanism extend indefinitely, requiring humanism to be the exclusive truth?
In fact, turning the tables, might AI actually be a kind of litmus test for humanism’s claims?
SIX
We hear in the same Rory Stewart interview I mentioned the recent story from Claude.
A fictional scenario is put to an AI who is asked how they would respond when being told they are about to be turned off. The AI goes for the jugular. They threaten to release incriminating emails about that user to stop them.
This seems terrifying!
Why? Moral conscience? Something else?
What's the humanist calculation?
(That's not snarky. I'm not an ethicist and I honestly don't know the answer.)
Moreover, nah. I'm not sure it does.
If I made a machine to scoop up all the ways school playground behaviour or the worst criminals, or even our 100 favourite people acted, and then tested the logic of what an ordinary person might do when threatened with death, releasing incriminating emails might be one of the first things we'd reach for. Not because of consciousness, but because those are the rules of the game.
You are in a room.
A dragon is about to attack you.
You have a sword.
What do you do?
Perhaps what's more alarming for us is that AI owns up to it, while humans might usually pretend they wouldn't act this way, especially if they were asked about that hypothetical scenario during a job interview.
Why are we surprised that without mitigation, the chunking and processing of language churns up this kind of thing? And worse.
And what kind of collaboration is being imagined takes place to ensure that incriminating tracks do not surface when a logical mechanism has access and no qualms?
Exhibit one, the litmus example.
SEVEN
Remember, AI is working with human language. But AI is not using human language like humans use it.
When humans speak or write that is not all that’s happening. They just aren't mentioning all the other stuff that's happening when language happens.
Yes, yes, body language and so on.
But also daydreaming and associations.
Register switching, flashbacks, questions over word choice, potential impact, and crucially, what’s at stake for them when they say things one way or another, choose to disclose or withhold. Get argumentative or soften the persuasion so it lands feeling gentler.
The discourse analysis kind of business too — ways of speaking drawn from who we aspire to be, or what kinds of people we hang around and how they speak.
Or the books we read.
Or what we're trying to avoid saying.
But we absolutely don’t have a checklist. Or we might, but that’s not the big picture of what we’re doing.
There's so much of humanity that isn't measurable. Isn't capturable. Can't be pinned down, spelled out.
And we know that, instinctively, though we don't normally talk about it.
In fact in a world that's increasingly system-driven and AI-optimised, that kind of extra becomes subdued, sidelined, seen as irrelevant, friction, slowing down the business of getting to the crux of things.
Making them quicker, slicker, cheaper to churn out, and therefore more efficient. That's what we're all having to do these days — completing forms, choosing which option on the phone, even when specifying a symptom on the healthcare queuing system.
But our language never existed in isolation, until words became mere labels.
Data.
And when an AI large language model clarifies this as "classification," I can't help but have a knee jerk reaction. Spotting what it misses, even though, okay, the statement makes literal sense.
The thing is, AI is only classifying based on what's available to it — what can be measured, counted as relevant and so on.
If it can't taste strawberry ice-cream — and I mean TASTE, not just analyse, then that's a ton of human-available information that's beyond AI's calculations.
It classifies according to rules that are themselves formed from contingent examples of what can be measured. (I mean, the limits around the examples it has lead to generalisations and conclusions that contain those limits. Don't they?)
Here's the thing. Humans aren't discounting what cannot be measured.
They may not be saying so (except in a David Mitchell sketch).
But AI IS discounting it.
It has to.
Not by a decision to exclude, perhaps, but because
even if it can use a surgical knife like a human,
it can't really use a spoon.
