Does word exist if we are not making it?

What does it take to make sense of the world around us?

I would spend days watching those vast, probabilistic engines of text,
churn out what they think is sense. Contrary to most people's
expectation, the current wave of generative AI doesn't simply challenge
our definition of intelligence. Rather, it forces us to confront a
fundamental failure of our own language.¹

Machines don't really perceive or sense. That much is clear when you
peer behind the curtain. What language models do is essentially map the
sprawling chaos of human writing, image pixels, and sound waves into
dimensions of vector embeddings. Every word, every token, is just a
point in an N-dimensional space, positioned by its statistical history
of the training data.

It's a beautiful, terrifying sleight-of-hand. The LLM processes these
smoothed, continuous vectors and projects them back into the hard,
discrete boundaries of human-readable symbols. This transition, from the
flowing vector space back to the rigid word, necessitates aliasing, a
loss that we have to accept every time we hit enter after typing some
poorly formatted strings in a chat box.

#Granting it a soul

Here's the strange, almost ritualistic behavior I've noticed in myself
and colleagues: when we question, complain, or curse the machine for
producing nonsensical garbage (or "hallucination," as the engineers call
it), we have already granted it a soul. We assume it can think, that it
understands our frustration at its failure to cohere with reality. It's
an almost innate, almost universal habit to treat the machine like
another person, attempting to parse words we ourselves did not make. And
we're left with the hollow feeling that a word without sense is like a
world without substance. It's the Turing Trap revisited: we
accept the output because we need the illusion of interlocution.

#Does the world exist if we're not watching it?

Can we formalize the world without these symbolic references? The
question always brings me back to the filmmaker Harun Farocki. Early in
his series Parallel II, where he explores the genesis of computer
graphics, he asks: "Does the world exist if we're not watching it?" Farocki was
asking about the computer's synthetic gaze, but the query is equally
potent for LLMs, which operate on a purely textual representation of the
world.

There's a history to this tension, i.e., the decades-long cognitive
science debate over mental imagery. In the 1970s, it was a philosophical
brawl. Was a mental image a literal internal picture (Kosslyn's
pictorial view)? Or was it just a descriptive, sentence-like
representation stored in a language of thought (Pylyshyn's
propositional view)?²

#What the machine sees instead

In the human mind, the pictorialists ultimately found support when
neuroscientists saw the visual cortex light up during imagination. We
seem to sense the world, then we name it. Sensation precedes symbolic
reduction. But the machines are the ultimate Pylyshnian dream. Their
reality is a pure propositional substrate. The vectors are just
compressed, high-dimensional descriptions. When a multi-modal AI
processes an image, it isn't "seeing" a cat, purring or stretching in
the sun. It's correlating the text vector for "cat" with the text vector
for "meow" and the numerical vector of RGB values that its training data
has already labeled "picture of a cat. An extremely efficient, un-human
way of cross-referencing.

Yet, our human impulse is to project our own experience onto it. We
think, "If it can describe the cat so vividly, it must see the cat." This
confidence that machines are replicating our mind is born of a profound
cognitive capture, where we confuse statistical fluency with ontological
understanding.

Machines don't really see; however, they are astonishingly proficient at
generating a lexicon of vision. They generate perfect captions and
stunning images based on verbal prompts. Words, in effect, compensate
for the machine's Achilles' heel: the lack of embodied,
continuous sensation. The curious parallel, though, is that we humans
share an almost blind, instinctive drive to reduce worlds into words.
It's an impulse toward epistemic closure. We want to pin down the
unruly, continuous flow of reality with the clean, discrete boundaries
of language.

#Zork's vocabulary

Think back to the old text adventure games like Zork I. The world
only existed through your input: 'TAKE SWORD,' 'GO NORTH.' If you typed,
'LOOK UP AT THE VAST, PURPLE SKY,' and the programmers hadn't included
'PURPLE SKY' as a recognized object or attribute, the world literally
ceased to exist in that moment. It was a digital existential crisis,
confined by the vocabulary. The analogy stands with modern AI. We
haven't changed much. We willingly subject ourselves to the machine's
"field of vision," modeled on our own neural complex, only to discover
this prosthetic vision is leaving us crippled.³

#The Vector-Symbolic Gap

The core theoretical obstacle, what I call the Vector-Symbolic Gap,
remains. Classical AI was logical. A word was a symbol with a fixed
referent. Modern LLM is probabilistic: 'king' is near 'queen' and 'man'
is near 'woman' in the embedding space because that relationship appears
often in the training data. This system knows the internal statistical
relations of symbols, but it has no necessary external relation to the
world. As Hubert Dreyfus argued decades ago, any system that lacks
embodiment and common sense (i.e., a situated background) is doomed to a
kind of abstract competence.

This is why the term "hallucination" is so misleading and, frankly,
dangerous. It's an anthropomorphic distraction. When an LLM confidently
asserts that "The Eiffel Tower was built by the ancient Romans,"
it is simply executing its function with perfect statistical integrity.
The vector space, having been trained on billions of tokens, has placed
"Eiffel Tower" and "ancient" (or "great wonder," "historical landmark")
in a proximity that is statistically strong enough to be selected,
despite the proposition being factually incorrect. It is a sampling
artifact, albeit however delusional it might sound. My proposal is to
replace that lazy term with "confabulatory interpolation". The
machines are performing a high-confidence, statistically coherent lie.
This is the fidelity of the statistical model, and it underscores the
true problem: the machine doesn't really care about truth.

#Bias in, bias out

The quality of the LLM is entirely determined by its corpus, the massive
training set. It is Borges' Library of Babel finally realized,
where all possible sentences are theoretically present, but the
algorithm's job is to efficiently retrieve the most likely ones.⁴ The old
GIGO principle ("Garbage In, Garbage Out") is far too simple. For LLMs,
it's closer to "Bias In, Bias Out (BIBO)." The data is overwhelmingly
biased towards English, corporate archives, popular internet text, and
easily digitized documents. This bias is amplified and structurally
embedded within the very geometry of the vector space.

#The difficult word

This operational text creates a crisis of the first-person perspective.
The LLM speaks with the voice of every author, and therefore no author.
We are losing the courage to pursue the singular path through language
that defines our individuality. We are being reduced to highly variable,
but ultimately replaceable, data points in the grand semantic field. The
true answer to Farocki's updated question is this: Words can exist
without human authorship, but they are hollow. They are statistically
perfect shells of meaning, lacking the intentionality and referential
grounding that only a situated, embodied, sensing, and fragile human
consciousness can provide.

We have to choose the difficult path. Our task is to use the
machine's very existence as a mirror to reaffirm the specific,
non-replicable value of human symbol-making. We must emphasize the
somatic, situated, and resistant nature of true human language. We must
push back against the ambient language by seeking out expressions that
are so specific, so improbable, that they defy the predictive model. We
must choose, once again, the painful, difficult word to save the
substance of our world.

#Footnotes

1. However unintentionally: that last sentence leans on a contrastive negation — "doesn't simply X. Rather, Y" — which happens to be one of the tics that makes LLM prose recognizable at a glance. ↩

2. Stephen Kosslyn's case for a genuinely pictorial mental imagery runs through Image and Mind (1980); Zenon Pylyshyn's propositional counter-argument is laid out in his 1973 paper "What the Mind's Eye Tells the Mind's Brain." The debate never fully resolved — it just moved to different substrates, which is more or less this essay's point. ↩

3. By the way, now even AI plays Zork — which says something about how far a purely textual world-model can get, and how far it can't. ↩

4. Jorge Luis Borges, "The Library of Babel" (1941): a library containing every possible 410-page combination of a fixed alphabet, almost all of it meaningless. The librarians' whole discipline is search, not creation — the same disproportion an LLM inherits from its corpus. ↩