The Accidental Semiotics Shuffle
Unavoidable problem of using loaded human terms as shorthand for AI mechanics, creating a semantic trap where narrow tech definitions are later mistaken for broad psychological claims
If you spend enough time listening to how we talk about Artificial Intelligence, you start to notice a weird vertigo. It’s the feeling of familiar words being used in deeply unfamiliar ways.
To understand why, we need to talk about semiotics.
Semiotics is the study of signs and symbols and how we use them to create meaning. It’s the invisible glue between a word (like “tree”) and the actual wooden thing outside your window. Normally, this connection is stable. But in AI, we are currently living through a frequent Semiotics Shuffle.
The Semiotics Shuffle is what happens when we take a messy, alien, purely mechanical process (like matrix multiplication in a GPU) and “shuffle” it under a familiar human label until it feels understandable. It’s rarely a con job; it’s usually just desperate shorthand. We have a new, terrifyingly complex thing, and we only have old, human-shaped words to describe it.
The Shuffle Mechanism
At its lowest level, an LLM is a frantic, hyper-complex pattern-matching game played in high-dimensional vector space. We don’t know if the LLM “knows” things the way you do; it has statistically probable continuations based on ingesting the entire internet (and more).
You can’t easily talk about that. It’s too alien.
So, we shuffle the definitions. We grab a nearby human concept—like “reasoning” or “understanding”—and accidentally smuggle in the human associations.
The machine can successfully complete a logical puzzle? We stamp it with “Reasoning.”
The machine can accurately predict the next word in a sentence about emotions? We stamp it with “Empathy.”
It’s a semantic shortcut. It saves us from having to say “statistically-derived token sequence that mimics human empathetic patterns” every time. But the confusion is often precisely in that gap.
Case Study: The “Introspection” Drift
A recent example popped up in a paper from Anthropic claiming to find evidence of “introspection” in their models.
“Introspection” is a heavy word. It’s deeply philosophical. It implies a “self” that is doing the looking (Or maybe doesn’t if you like some forms of meditation, things are complicated on the human-only side alone!) Anthropic found a fascinating mechanism: a pattern of neuron activations that lights up when the model is about to output something conflicting with its training. It’s a “betrayal” monitor.
Calling this “introspection” isn’t malicious, but it is a semantic leap. A critique of this paper (captured in the image below) points out the “methodological amphigory” here.
The researchers admit that genuine internal representation is “difficult to demonstrate directly.” So, they settled for indirect evidence (asking the model yes/no questions). They needed a word for what they found. “Introspection” was the closest human analog, so they used it. The concept was shuffled from a mechanical reality into a psychological trait, almost by accident, because what else were they going to call it?
The Citation Trap
This shuffle becomes genuinely dangerous in the citation chain.
Academic papers often define their terms very carefully in section 2, paragraph 4. They might say: “For the purposes of this study, we define ‘understanding’ as the ability to correctly answer 80% of questions in Benchmark X.” That’s a perfectly valid, narrow, technical definition.
The problem happens when that paper gets cited. The nuance evaporates instantly. The next person just says, “Smith et al. (2024) demonstrated that models now possess understanding.” They have dropped the narrow technical definition and reverted to the fuzzy, broad, magical human definition of the word.
It’s a bait-and-switch that no one intended to pull. The rigorous definition in the original paper provides “cover” for the sweeping claims made later by people who only read the abstract. We use the credibility of the technical achievement to smuggle in the implications of the human concept. It ends up being a motte/bailey fallacy.
The Inevitability of the Shuffle
We almost have to do this. We are hairless apes who only understand new things by slamming them together with old things.
We called early cars “horseless carriages.”
We talk about “computer viruses” (even though they lack biological life).
If AI researchers used technically accurate terms for everything, no one could read their papers. We need metaphors to bridge the gap between our analog brains and these digital realities.
The Danger is Forgetting it’s a Metaphor
The problem isn’t the metaphor. The problem is forgetting it is a metaphor. When we perform the Semiotics Shuffle enough times, we stop seeing the mechanics and only see the label. We start treating the map as the territory.
We don’t need to stop using these words—they are useful pointers. But we need better semantic hygiene. We need to remember that when we say an AI is “hallucinating,” it’s not having a psychedelic experience; it’s having a statistical error.