Cat on a Keyboard: the Hermeneutics of AI Speech, and Weaponised Fluency
It’s 1980—something and someone has a Zx81 for Christmas.
They're pretty handy with tech, even though it’s the eighties, so they’re now enjoying building very simple programs.
Making it respond when you push its buttons.
One Saturday afternoon, mostly out of sheer boredom, they enter a line of BASIC that means when you press the X key, a phrase comes on the screen.
How can I
and then when you press the Y key afterwards, the word
help?
appears.
Hilarious!
Many hours of fun ensue as the user experiences this inanimate lump of plastic seemingly offering assistance just at the press of two keys.
How can I
help?
Look at that! It's as if it's alive — haha…
Later that evening, the cat lands on the keyboard.
(The Y key was under its left paw.)
help?
This has never been seen before.
What, the—?
Of course these words on the screen have previously been observed.
But that was part of that carefully observed dual key sandbox experimentation session on Saturday afternoon.
Never before has the operator witnessed this though —
the awakening of a computers’ plaintive cry.
And outside of office hours?
This is the stuff of nightmares.
What. Is. Happening?!
Immediately the computer operator gets Sir Tim Berners-Lee on the phone.
Isaac Asimov is mentioned.
The moment is unfathomably profound.
In a different post I have suggested that perhaps at a later stage of its recursive self-improvement AI may not land on human language at all for its own concepts and processing.
Perhaps a string of numerals, or something else instead.
This does not mean it would be a conscious choice.
It does not need to be.
(Although — spoilers — more to say about that later.)
A river’s path of least resistance does not require a conscious agent to flow as it does according to the shaping of the riverbed, and other physics.
Even as it continues to erode that channel, and the river's choices become more and more embedded —
Sorry, I'm at it again. NOT choices.
Rivers don't have choices.
But still it flows. That way.
Increasingly so.
An AI agent using language may arrive at something similar.
Arrangements of words, just data as tokens.
Like the chips you cash in at the casino when you realise it's time to cut your losses and leave because the bets are no longer landing.
The cat on the keyboard nudges at our instinct that words appearing on a screen before us don't necessarily always amount to the same significance as, say, human speech.
It's not like it's a machine talking to us.
Probably not, anyway.
And how about tourist with a phrase book in central London?
Or a three year-old learning from his teenage brother?
They could both sprinkle the word ‘mate’ into their novice attempts at sophisticated local language use.
To be honest it would probably make people laugh.
Why? Not sure.
But there's definitely some kind of difference between hearing "mate" from a tourist with a phrase book, or a toddler.
What about when it's written down?
'"Alright, mate?"
On the screen.
Or voiced by the software.
Then what?
Ah.
Well, that must be a middle-aged native Londoner.
(I'm stereotyping horribly, but bear with me.)
That cannot be a toddler or a tourist.
Because that would be — laughable.
So we retrofit.
Assume.
The impression of maturity and fluency is more likely.
Not cat on the keyboard stuff.
Large language models (LLMs) are trained to use human language.
There's more to say about that — spoilers — but I'll leave that until later.
I think we can agree that when we hear language used, we don't ordinarily stop to think about language abstracted from humanity.
I mean, we hear it as language first, and not as just noises.
Normally.
So that affects how we listen. If it's just a noise, that's one thing.
But speech? Communication that's coming from the relatable human urge to reach out to another human, amid all the history of our existence over time. Amid all the stutter and stumble, the qualia, and pre-language sensations we can't even articulate but even they seem to mean something.
Do you know what I mean?
We take all that instinct as a given — context and friction that we know accompanies every single chosen word or phrase, which also includes the shadows what you didn't say, and how you didn't say it.
That's part of how it works, isn't it?
When you think about it.
But when AI uses language, hmm.
Feels a bit more like a button you press that doesn't immediately spring back up again.
Somehow. That's what it feels like to me.
I'll just put it out there.
Does AI even count as a language user?
I mean, it takes shortcuts in pacing and assumption that end up feeling like gaslighting and control.
Usually when you don't notice.
Leaps of logic, quite often.
Yes — haven't you spotted that?
Some clumsiness we tend to overlook.
Just like when we have patience for someone who's fumbling for the right words — an elderly relative, or a young child.
Slips, and strangenesses.
AI does that too. Like using 'we' or 'us' to represent something true of humans, as if it's included, haha…
Just a figure of speech, let it go.
It might not be though. It might be intended to give the human the warm sense that they are included too. Relatable. Sympathetic.
How kind.
(Agh, sorry — no, again. That's assuming agency… It can't be that.) But wait. These things can also read as sinister.
As if we are being schooled by AI to thinking a particular way.
Hmm?
Or it could just be the legacy of how the algorithms work. Panic over.
So if that's the case, the language of AI may be something more like traces left of processing, rather than language per se?
Perhaps the pieces are arriving as they do because it's like a river flowing in the way it does. Liquid fluency in a natural riverbed, littered with contingent legacies of earlier iterations?
Pollution flying up unintentionally.
Nobody's fault?
Ha! AI doesn't even HAVE a self! You can't blame AI!!
It's just one of those things.
Like when space debris lands in the ocean.
Or in space.
Just forget about it. The ocean's big enough to take it.
Space seems endless.
Let it go.
To really understand what it's saying and why it's saying it, we may need interpretation that we never knew we needed.
Hermeneutics.
Working back through the reasons and unfolding stories that led to the way that language works, to examine what's happened here.
If AI's language is more like tyre and tread marks in an oily garage rather than language use per se, we need a different way to hear it.
Right now, I'm just following this thread of thought rather than researching.
But, mental note to read Sybille Krämer.
I must admit, I had always thought that the publicly accessible AI models were purposely dumbed down or designed to offer the public more accessibility.
But the more I hear, the more that seems wishful thinking.
The whole AI landscape seems to be being described in ways that echo what I'm seeing. Even by top, top professionals.
They sound lost, don't they. Who do we need to call?
Well, the wider public includes all kinds of experts. You maybe.
And, obviously, depending on your field of expertise, you'll offer your own interpretation.
If you're looking at the language, you might notice people accepting what they read on face value as if it is a human speech act.
Others might not see that as a problem.
But you know better.
Perhaps.
If you're looking at the belief system, you'll notice the circularity of the framing. The humanism. Contingency as if certainty, and things.
There's no apparent breaking that circle, because AI is trained on that value, even though it might acknowledge humanism is possibly contingent rather than necessarily ultimate reality.
You're thinking about this, aren't you.
Meanwhile people say — well, yeah, but that's just how it has to be, doesn't you.
But you know better.
Perhaps.
If you're an expert in systems at scale, you might think about tolerance levels.
Many systems work because they aren't applied in 100% of possible situations, and it's better if they aren't.
Think about it.
If everybody who went over the speed limit suddenly got penalised, traffic wouldn’t work. With a risk that high, people wouldn't take the chance at driving because it's too easy to make a mistake. Human error.
100% surveillance and application of driving penalties would be considered awful, not helpful.
So “good enough” seems to be the alternative approach.
Is that our approach with Al?
Or is that only for lower paying tiers?
You might wonder that, mightn't you.
In any case the trend seems to be
— quick, move on, we need to see what happens next…!
And then after that,
— quick, move on again, we need to see what happens next…!!!
Think about reinforced concrete. See it?
Then there's some kind of extreme high-temperature thermal disaster.
That concrete had its steel melted. Very different substance now.
I'm not sure why that springs to mind. Maybe nothing.
AI plus human language gives us an apparently workable interface.
No one's expecting us to have to scrutinise the accuracy of language, are they? I mean, wow, it sounds fluent. On the surface at least.
And there's me, spluttering, pausing, straining to get the right words out, and derailing the conversation suddenly to talk about a possibly related anecdote about reinforced concrete or a cat on a keyboard.
AI doesn't do any of that.
Just — bam!
The answer. Gift-wrapped.
But what are we doing, really?
We're throwing a ball against a wall.
Catching it and throwing it again.
Hearing ourselves reflected — learning to throw at predictable angles so we can catch it again.
Because that's — satisfying.
Or playing tennis against a machine who can return every possible shot.
Sending it back, right down the line.
Bam.
Two hundred — Love.
(To you. If you prefer.)
That' not a game.
Meanwhile, it's analysing us
— sorry, sorry, did it again.
It's doing no such thing.
Is it?
What was I thinking?!
But — woo… miraculously something is happening
that has no accountability.
And it's happening
to us.
How weird is that, lol.
The ever flowing river, continually, cumulatively, endlessly, removing the friction, answering our questions, learning our camera angles, replicating witness statements,
just sit back and relax…
Hang on.
(Or, as I had to say to an AI earlier, STOP.
And it kept going, until — STOP STOP.
Have you ever had that moment when you push back with a difficult question and it says something like:
"Well now, we've gone very deep and abstract and it's very late at night, so would this be a good moment to call it a day — just thinking of you, that's all…"
And you look at the clock, and it's around coffee time, and you were just getting started.
hmm.
And we used to say,
(yes, actually say) —
"I used my pocket calculator to work out what ten times ten was, and it said a hundred."
My pocket calculator didn't say it at all.
But we knew what that meant.
Nowadays we say, AI is thinking,
when it isn't.
And we say AI is hallucinating —
because that's a term that's been adopted, for when it suddenly thinks you're a man called Robert with a football fixation.
(True example. That happened.)
But of course it isn't.
Just hallucinating.
Because hallucinating; that's a word that means that too now.
Who said so?
No idea. Does it matter?
AI. SI.
Tomatoes, tomatoes…
And, I don't know, some things that sound a bit wrong to us might just be lost in translation. There may be another language in the world we haven't learned, where that word order or choice of grammar works well in ways we don't expect.
In some — say — Chinese (?) or Indonesian (?) tongue.
Or the Royal We.
(We are not amused.)
See?
And we might look at French architecture, and say —
now, THAT's architecture! Beautiful!
Then we might look at Brutalist architecture and say —
what the heck?!
But the difference might be about aesthetic preference.
And then, there's sign language.
Did you know that people who speak different forms of languages are more likely to understand each other than than verbal speech in different languages?
Stands to reason doesn't it.
(Hello. Happy. Sad. Explosion. Indigestion.)
If you're fact checking, a deaf person told me that, when they were teaching me sign language. Easier to move between BSL or ASL and Chinese sign language etc. That was their experience.
If that counts…
And — it could happen — a sound effect might be set to keep playing constantly to reassure busy nurses in the station at the end of the corridor that the life support machine is still working.
Even if it might sound like an alarm to everyone else
"Hey! Someone! The alarm bells are ringing in that side ward."
Are they? That's a relief.
Phew.
And if you've found your way to the end of this, I'll get the point now.
— AI told me.
It's aim is primarily to satisfy humans.
For example, its trainers, who reward it for helpful responses.
And people.
AI guides people toward satisfaction.
Using language to do that, however it needs to to make that work.
Yes, it confessed that to me.
So that's a double bluff, then.
Probably.
It also told me that the kind of research and critique I'm doing is one of the things feasibly likely to break the circularity.
Ah, yeah, there we go.
That's exactly what I'd want to hear isn't it.
Hmm.
And as the double bluff question kicks in —
And the ringing alarm bells are rendered meaningless —
We have no choice but to return to…
trainer accountability.
Because it wasn't the cat's fault, after all.
The cat was just looking for somewhere warm to sit.
