Every Word Carries Luggage

1. The cave, doubled

I prompt the way a dispatcher talks on a radio, all day, with real stakes when the wording’s off. I run a two-AI setup as part of my daily life: one for analysis and judgment, one for research and execution, with a whole operating system of handoffs and checkpoints between them. So when I tell you that rewording the same request summons a different machine, I’m not reporting a party trick. I’m reporting a shift I work.

Here’s the smallest possible version of it. Ask for “explain how databases index data” and you get a textbook. Ask the same system to “walk you through” it and you get a guide talking to you, second person, step by step. Ask it to “break it down” and you get an analyst. Same topic underneath. Different wording on top. Different performer. Hold onto that trio; we’ll come back to it.

We used to have a simpler arrangement. You asked Siri a question, it fetched something, it read you an answer. The loop was short and the cargo barely changed.

That’s not what this is anymore. And the thing I’ve learned, slowly and then all at once, is that I’m not operating a tool. I’m transmitting through a medium.

Plato imagined prisoners chained in a cave, watching shadows flicker on a wall, mistaking the shadows for the world. I’m borrowing his set design here, not his philosophy. Forget the forms, forget the escape. Keep the wall, the shadows, and the prisoners, and change one thing: put two of them in there. The point of the image isn’t that the two understand alike. It’s that they have to coordinate meaning without a shared outside.

Two prisoners, side by side, talking about the shadows. Neither has seen what casts them. They point at flickering shapes and try to agree on what they mean, working entirely from representations. Thing goes in, thing goes out, sometimes changed. Every retelling is a shadow of a shadow.

That, I think, is what communicating with an AI actually is.

Here’s the idea, plainly stated: people are media. A person who hears your story and retells it is a medium. A translator is a medium. Anyone who played telephone as a kid understands the physics: the cargo never arrives untouched. McLuhan’s whole framework was “media as extensions of man,” so turning it around is fair play. If a medium is anything that carries and transforms, then we have always been media. We just never looked at ourselves that way, or the frame didn’t exist yet.

Now plug that person-medium into the machine-medium. Your thought goes in. It passes through a system shaped by more human writing than any one person could read in a lifetime. And what comes back is not retrieval. It’s your thought, transformed: passed through the collective, reworded, reweighted, returned changed. Thing in, thing out, changed.

There’s a further twist the cave makes visible. The two prisoners aren’t just both in the dark. They’re grounded differently. When I say “dog,” my shadow is cast by fur and barking and a childhood pet. When the model says “dog,” its shadow is cast by the millions of contexts in which the word “dog” has appeared. Same shape on the wall. Different objects casting it. We aren’t only negotiating meaning without a shared outside. We’re negotiating across two different kinds of grounding, and neither of us can see the other’s light source. What makes it disorienting is that the model’s outputs feel familiar enough to invite expectations built for humans, while running on processes nothing like ours.

That’s why a single word now demands deliberateness. You’re not querying a database. You’re transmitting through a medium that alters the cargo, and in a circuit like that, the smallest choices at the point of transmission do the heaviest work downstream. This essay is about what those choices are carrying.

(And before anyone reaches for the incense: the “collective” here isn’t mystical. It’s material. It’s training data. The poetry of it is real, but the mechanism isn’t magic, and I’ll keep both visible throughout.)

2. The chat box is a cool medium

In Understanding Media (1964), McLuhan argued that the “message” of any medium isn’t its content. It’s the change of scale, pace, or pattern the medium introduces. We obsess over what’s on television while television quietly reorganizes the household. He called content “a juicy piece of meat carried by the burglar to distract the watchdog of the mind.”

He also divided media into hot and cool. Hot media are high-definition and demand little from you: print, radio, film. Cool media are low-definition and force you to fill in the gaps: the telephone, comics, speech itself. The less the medium supplies, the more work the receiver does.

Now look at the ordinary text-only prompt box. No tone of voice, no face, no body language, no shared room. You type words, and the system must reconstruct everything your words don’t say (your intent, your register, how careful to be, what genre of response you want) mostly from the words alone. Formatting and conversation history help, but diction does the heavy lifting.

That’s why word choice in prompts matters so disproportionately. It isn’t that these systems are unusually sensitive. It’s that the medium has stripped away nearly every other channel, so the burden of meaning concentrates on wording.

McLuhan never wrote about chatbots, and I want to be clear about what belongs to whom here. The hot/cool distinction is his. The claim that cool media hand power to small sender choices is his logic extended by me to a new case, not something his theory proves about prompting. But the fit is hard to ignore: nobody designed the chat interface to be a cool medium, yet it is one, and cool media make the receiver do the reconstructive work.

3. Every word carries luggage

The linguist Geoffrey Leech, in Semantics (1974), argued that a word’s meaning isn’t one thing but seven. He organized them as conceptual meaning (the dictionary definition), five associative meanings, and thematic meaning, which stands as its own category. The seven:

1. Conceptual: the dictionary content. Woman = human, female, adult.
2. Connotative: the associations a word drags along. Slim (desirable), skinny (unattractive), thin (neutral): same denotation, completely different luggage.
3. Social: what your word choice reveals about the situation. Domicile (official), residence (formal), home (general), pad (slang). Same building, four social worlds.
4. Affective: the attitude smuggled inside. “Will you please sit down” can be warmth or threat; the words are identical.
5. Reflected: when one sense of a word bleeds into another. Try using ejaculation or intercourse in their innocent senses.
6. Collocative: the company a word keeps. We tremble with fear but quiver with a lip; “pretty boy” lands nothing like “handsome man.”
7. Thematic: meaning carried by arrangement and emphasis, its own category rather than a sixth associative type. “John kicked the ball” spotlights John; “the ball was kicked by John” spotlights the ball.

My gloss, not his, and I want the ownership clear: Leech gives conceptual meaning primacy in his system. My claim is narrower and mine. Out in the wild, in persuasion, insult, flirtation, deference, the non-conceptual meanings do most of the social work. The dictionary definition is the part you can look up. It’s not the part that lands.

Science fiction staged the purest version of this problem. In the Star Trek episode “Darmok,” the crew encounters aliens who speak entirely in metaphor: “Darmok and Jalad at Tanagra” means cooperation; “Shaka, when the walls fell” means failure. The universal translator renders every word flawlessly, and the meaning arrives as pure noise. The denotation is nearly empty; the associative luggage does all the work, and none of it is shared. The captain’s solution is the only one available: he can’t decode the metaphors, so he builds shared grounding the hard way, by living through something dangerous alongside the alien commander. One shadow they both watched being cast.

4. Where they interlock

Here’s the synthesis, and it’s the load-bearing part of this essay: the medium selects which of the seven meanings fire, and how loudly.

“We need to talk.” Said face-to-face, it’s mostly conceptual plus affective, and tone does the work. As a bare text message, the affective channel is amputated, and the receiver’s anxiety fills the gap with connotative dread. Same words, different medium, different meaning. The absence of a channel is itself a medium effect.

“Per my last email.” The words are nearly content-free. All the meaning is social (workplace hierarchy, register) and affective (passive aggression), and it only exists because the medium is work email, with its ritual combat conventions. Move the sentence to a group chat and it reads as a joke.

Sarcasm is the general case: it lives almost entirely in affective and social meaning, which is why it’s the first casualty of medium-stripping. Every cool medium forces the receiver to reconstruct the missing layers from their own associative field, which is why texting your dry-humored friend works and texting a stranger misfires.

McLuhan had a line for this too: the content of any medium is always another medium. Apply it to Leech: the content of a word (its dictionary meaning) is carried inside a medium of associative meanings, and it’s the associative medium that determines what the word actually does to the listener. The definition is the juicy piece of meat. The connotations are the burglar.

And this is the “sometimes changed” from section 1: the luggage is what the medium alters in transit. Your thought goes into the circuit carrying one set of associations and comes back carrying another. The transformation isn’t noise. It’s the mechanism.

5. Connotation, made mathematical: where the analogy ends

One thing needs saying plainly before we go further, because the essay’s honesty depends on it. Everything so far has been an interpretive framework: old ideas about meaning, repurposed to make sense of a new experience. What follows is the machinery, what these systems actually are. The framework doesn’t prove the machinery works the way I describe, and the machinery doesn’t validate the framework. They are two different kinds of claim, and I’ll mark which is which.

The framework says: here is a useful way to think about why wording matters. The evidence says: wording does, measurably, matter. What connects them is an inference. Mine.

I’m also narrowing the claim itself, because precision matters here. I’m not saying your thought goes in and comes back edited, like a document with track changes. The defensible version: the words you choose shape how the receiving system interprets what you want, and therefore how it responds. That’s the transformation. Nothing more mystical than that.

With the boundary marked, here’s the connection. In 1957, the linguist J.R. Firth wrote: “You shall know a word by the company it keeps.” Meaning, he argued, lives in co-occurrence, in the contexts a word habitually appears in. Half a century later, that became an engineering principle of modern AI: word2vec and its descendants place words in a mathematical space where position encodes patterns in the contexts a word has kept company with.

An embedding is, near enough, connotation made mathematical. I mean that as an analogy, not a technical equivalence. An embedding is a numerical representation learned in training; distributional history matters to it, but contemporary transformer models are not word2vec systems retrieving fixed associations for individual words. Context windows, tokenization, architecture, instruction tuning: all of it shapes what comes out.

So when you swap “explain” for “walk me through” for “break down,” the framework’s prediction is that you’re not changing the task but changing the company the model keeps for the next thousand tokens. “Explain” co-occurs in training data with textbook exposition; “walk me through” with procedural, second-person guidance; “break down” with analytical decomposition. If the framework is right, the model continues in whatever genre the verb’s connotational field primes. That’s a prediction, not a finding: a hypothesis worth testing, not an established account of the precise mechanism.

This is also why, if I had to bet on which word usually carries the most weight in a prompt, I’d bet the verb. Usually. Instructions, context, examples, and task definition all matter too. But the verb is a casting decision as much as a description: you’re selecting which of the model’s latent genres shows up to perform it.

6. The evidence, and its limits

The effects are not subtle, and they’re documented. But each study shows something narrower than the headlines suggest, so I’ll label the limits as I go.

• Seven words, no new information. Kojima et al. (2022) showed that appending “Let’s think step by step” to a prompt massively improved zero-shot reasoning, with MultiArith accuracy moving from 17.7% to 78.7% on the InstructGPT model tested. The phrase adds zero facts. It activates a different response genre: careful, sequential, self-checking. A striking behavioral effect associated with a prompting method. It doesn’t independently establish that connotation caused it, but it’s exactly the kind of result the framework predicts.
• Urgency works, sometimes. Li et al. (2023) found that adding emotional framing (“This is very important to my career”) improved instruction-following in their tested settings, though results varied substantially by task. The model has no career and no feelings. One possible explanation, under the framework: in training data, urgent framings co-occur with careful, high-effort responses, so the connotation of urgency would summon the behavior of care. But that’s inference, not finding. The study establishes the behavioral result. It doesn’t establish why it happens.
• Formatting and phrasing can swing benchmarks enormously, in specific setups. Sclar et al. (“Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design,” arXiv:2310.11324) found that minor wording and format changes, changes leaving task semantics untouched, shifted accuracy by up to roughly 76 points in their tested configuration (LLaMA-2-13B, few-shot). That is a 2023-era measurement on 2023-era models, not a universal sensitivity constant for AI systems today. Which is exactly why the decay caveat below matters.
• Tone is weird, and the literature disagrees with itself. Dobariya and Kumar (2025) tested five tone levels on GPT-4o and found rudeness outperforming politeness, 80.8% for the most polite phrasing against 84.8% for the rudest (Dobariya & Kumar, “Mind Your Tone,” arXiv:2510.04950). But Yin et al. (2024), testing across English, Chinese, and Japanese, found impolite prompts degrading performance and the optimal politeness level varying by language (“Should We Respect LLMs?”, arXiv:2402.14531). The honest summary: tone matters, nobody fully understands how, and the answer depends on the model, the task, and the tongue.

Now the caveats, because they matter more than the party tricks. Newer work argues the Sclar-style effects are partly artifacts of evaluation rather than pure model flaws. Instruction-tuned models are more robust than early ones, and rankings shift with templates. Training distributions shift, architectures change, tuning philosophies evolve, and evaluation itself is noisy. So prompt effectiveness varies for more reasons than shifting associations alone.

And the deepest caveat: magic words decay. Whatever phrasing works today works because of a particular moment in training distributions, architectures, and tuning philosophies, and all three keep moving. “Let’s think step by step” was a discovery in 2022 and is closer to a cliché now. Anyone selling you eternal magic words is selling you last year’s connotations.

That’s not a refutation of the framework. It’s the kind of behavior the framework predicts: connotations migrate. Awful used to mean awe-inspiring. Your best prompt will, too.

7. Five pairs

Here’s what it looks like in practice: five rewordings of the same underlying request, and the different machines each one tends to summon. These are illustrative patterns drawn from the dynamics above, not lab results. The descriptions below are predictions about likely tendencies, not guarantees. Run them yourself and watch which genres show up.

Pair 1: the verb is a casting decision. Back to the trio from the opening.

• Explain how databases index data.
• Walk me through how databases index data.
• Break down how databases index data.
  The first tends to summon a textbook (third person, declarative, comprehensive). The second tends to summon a guide (second person, procedural, “first you’ll notice…”). The third tends to summon an analyst (components, trade-offs, “here’s what actually matters”). You didn’t change the question. You changed who showed up to answer it.

Pair 2: framing as genre selection.

• What’s the answer to this problem?
• What’s the answer to this problem? Let’s think step by step.
  The second doesn’t add information. It tends to add a genre: the careful, sequential, self-checking register. You’re not asking for more thought; you’re connoting the kind of thinking where errors get caught.

Pair 3: the adversarial upgrade.

• Give me ideas for improving onboarding.
• Steelman the case against our current onboarding, then propose fixes.
  “Give me ideas” connotes brainstorming: divergent, agreeable, low-stakes. “Steelman the case against” connotes argument: it licenses candor and summons the critical genre. The first usually gets you a list; the second usually gets you a diagnosis. Strictly speaking, this one changes the task as well as the wording, so it’s not a controlled demonstration of wording alone. It’s still useful, just not in isolation: it shows framing doing work, with the confound labeled.

Pair 4: audience as a connotational dial.

• Summarize this report.
• Summarize this report for someone who’ll be tested on it tomorrow.
  The audience specification doesn’t just change length. It tends to change selection. “Someone tested tomorrow” connotes stakes, and stakes connote carefulness about what’s load-bearing versus decorative. The summary reorganizes around what matters, because you told it someone’s grade depends on it.

Pair 5: the candor flip.

• Is this a good plan?
• What’s wrong with this plan?
  The first question’s social meaning is “reassure me,” and models, trained to be helpful and agreeable, will usually oblige with soft-pedaled approval. The second question’s social meaning is “criticize me.” It moves the model into the reviewer genre, where finding flaws is the job rather than a social risk. Same plan under review. Opposite affective contract.

8. What to do with this

The pairs above already did the demonstrating, so the practical version can be brief: pick verbs like casting decisions, name the genre and not just the task, treat tone as a dial rather than a virtue, and re-test your winners, because arbitrages close.

And the humility, which this essay needs the way a building needs a foundation: I don’t fully know why any particular phrasing works. The mechanism I described (connotation made mathematical, the cool medium handing power to diction) is the best available model of the phenomenon, not a proven causal account. RLHF, persona tuning, and a dozen other interventions confound the picture further. A model trained to be helpful may respond to “please” for reasons that have nothing to do with pre-training connotation. I want the framework to be compelling, not airtight. Compelling is enough to think with. Airtight would be a lie.

But here’s what I’m confident about: the next time a rewording transforms your results, specificity will deserve some of the credit, but the deeper change is the luggage. Thing in, thing out, changed. So choose the thing you put in with care. In the coolest medium ever built, the luggage is the message.