Big Rooms Drown Small Words

01The room

Earlier today I called it dilution. In Not Is a Nudge, the measurements showed that “not” turns a sentence only a little, because it is a small share of everything the sentence carries. Here is a better picture of the same idea: a room full of voices.

When one of these search systems gives a passage its address, it listens to every word in the passage and settles on one overall impression. For the two models measured in this series, that is not a figure of speech. Each word-piece gets its own vector, and the passage’s address is the plain average of all of them. Every word is a voice in the room, and the address is what the room sounds like from the doorway.

In a small room, one voice carries. Say “not” in a room of eight words and the room sounds different. In a big room, the same voice is still speaking, just as clearly, and the room barely changes its tone. Nothing was deleted. It was outnumbered.

02The hinge words

Linguists divide words into two kinds. Content words, the nouns, verbs and adjectives, are an open class: new ones are coined every year. Function words are a closed class: a small, fixed set of short, very common words that hold the grammar together. The machine hears content words loudly, because content is what makes a passage about something.

But a small set of those function words, and a few near relatives, do something outsized. They do not add a topic. They bend the one that is there. I call them hinge words.

• not
• no
• never
• unless
• except
• only
• before
• after
• all
• some
• may
• must
• until
• notwithstanding

Here is the strange history. For most of the life of computer search, these words were thrown out on purpose. In 1958, Hans Peter Luhn at IBM proposed that a computer indexing text should first discard the most common words by checking them against a stored list. Within a few years such lists were called stop lists. The classic ones include “not,” “only,” “all,” “before” and “after,” and most include “except.” For decades, the standard first step in search was to delete exactly the words a contract turns on.

Modern AI search does not do that. It keeps every word. But it inherited the same imbalance by a different route: the hinge word is now kept, and then averaged into a room full of content.

03What each hinge does, alone

Start with the smallest room: one sentence. Eleven ordinary sentences were each changed by a single hinge and measured against twenty unrelated sentences. Every time, in both models, the changed sentence was the original’s nearest neighbor out of 41 candidates, however completely the hinge had changed the meaning.

Figure 01

How far one hinge word turns a single sentence

• “The tenant pays the landlord” → “The landlord pays the tenant“6°
• “Pay the contractor before / after the work is finished”11°
• “The dog bit the man” → “The man bit the dog”12°
• “Refundable within 30 days” → “within 3 days”13°
• “This medication is not safe during pregnancy”14°
• “Only the buyer may cancel the contract”18°
• “The seller may / must repair the roof”19°
• “Always / Never turn off the gas before leaving”20°
• “All / Some of the units passed inspection”22°
• “Tenants may keep a dog unless the landlord objects in writing”26°
• “The warranty covers all parts except the battery”27°
• Two sentences on unrelated subjects~86°

Look at the top of the chart. Swap who pays whom, and the sentence moves six degrees. Every word is the same; only the order changed, and the order is the whole meaning. Researchers testing image-and-caption models found the same thing on a larger scale: shown two captions built from identical words in different orders, the models did little better than chance at matching them to the right pictures, and on some measures worse.

04Now make the room bigger

Nobody searches single sentences. Real systems cut documents into passages, often a paragraph or more, and give each passage one address. So here is the experiment that matters. Take the boat-lift sentence and its negation, and surround both with the same ordinary contract language: closing location, prorated taxes, the termite letter, the final walkthrough. Add one sentence at a time, and measure how far the “not” still turns the whole passage.

Figure 02

The same “not,” heard in a growing room

• 1 sentence (8 words)18.2°31.0°
• 2 sentences (20 words)9.9°11.0°
• 4 sentences (43 words)5.0°3.6°
• 7 sentences (73 words)3.3°2.5°
• 13 sentences (132 words)3.6°1.8°

• all-MiniLM-L6-v2 (384 dimensions)
• all-mpnet-base-v2 (768 dimensions)

By two sentences, the “not” has lost about half its voice. By four, most of it. By a paragraph, the passage that says the seller keeps the boat lift and the passage that says the seller leaves it are almost the same address. In the larger model, the two are under two degrees apart. For comparison, the faithful paraphrase in the last piece sat 35 degrees from the original.

The “not” never stopped speaking. The room just got louder.

05The turn: the room is a choice

Here is what I did not expect. The size of the room is not a law of nature. Somebody picks it. Every system that searches documents has to decide how to cut them into passages before giving each one an address, and that decision sets the room size for every hinge word in the document.

And it is a genuine trade. Big rooms give context: the address of a whole paragraph knows it is about a boat lift in a residential sale in a particular spot. Small rooms let the hinge word be heard. But small rooms have their own failure, and it is the nastiest one. Hinge words often act at a distance. The rule is in one paragraph: “Tenants may keep one dog.” The exception is three pages later: “Notwithstanding anything in this lease, no pets are permitted in units with carpet.” Cut the lease into small rooms and the exception is loud, but it is in a different room from the rule it overturns.

People fight over hinges too, of course. In 2017, a federal appeals court found a Maine overtime law ambiguous because of a missing comma in a list of exempt activities. The dairy drivers who brought the case settled the next year for five million dollars. Lawyers have always known that the smallest marks carry the largest consequences. What is new is a reader that, by construction, hears them least.

06What I expect to see

1. Documents get a second index, just for hinges.

   Serious document systems will track the hinge words separately: every “not,” “unless,” “except,” “only,” “notwithstanding,” date and number, with a pointer to the rule each one bends. Meaning finds the passage. The hinge index says what can overturn it.

2. Room size becomes a setting people care about.

   For contracts, medication and safety instructions, systems will read in small rooms, clause by clause. For everything else, big rooms are fine. “What size does it read in?” will become a question a careful professional asks about any AI tool.

3. Good answers lead with the exception.

   The assistants people trust will answer “Yes, unless…” and “No, except…” up front, because the exception is exactly what the machine is structurally worst at hearing, and the one thing the person most needs to know.

4. Drafting moves the hinge next to the rule.

   Writers who know machines will read their documents will put the exception in the same sentence as the rule it modifies, and “notwithstanding anything elsewhere in this agreement” will come to be seen as a hazard, not a convenience.

5. Human reading concentrates on the hinges.

   The expert’s job shifts from reading everything to reading the handful of words the machine flags as able to flip the answer. Ten pages of contract become the dozen sentences that contain a hinge.

07Why I keep coming back to this

The realtors I have been working with this month live inside documents made of hinges. A purchase agreement is mostly boilerplate, and the fights come from a few words in it: an “unless” in a contingency, an “except” in the list of what conveys, a date before which something must happen. A transaction manager that finds the right contract by meaning and then misses the hinge inside it has done the easy part and skipped the part that costs money.

This is the fourth piece in a row on one idea. High dimensions give AI its strange abilities. Meaning becomes an address. “Not” is only a nudge at that address. And now: the nudge gets smaller the more the machine reads at once. Same geometry, all the way down.

08The ledger

Already true

In widely used embedding models, one hinge word moves a sentence far less than a change of subject does, and its effect shrinks steeply as the passage grows. Both models measured here build a passage’s address by averaging its word-pieces, so the room picture is literal for them. Image-and-caption models have been shown to ignore word order at chance level, and standard search embeddings do worse than chance at telling a passage from its negation.

What has to happen

For the predictions to hold, search has to keep summarizing passages into single addresses, and documents have to keep being cut into passages of a paragraph or more. The second reading step, clause by clause, has to stay cheap enough to run where the stakes are.

Where I am probably wrong

These are two small open models and a handful of sentences; commercial models may pool differently and train harder on exactly these cases, and brief fine-tuning is known to fix many quantifier failures. The room is a simplification: each word’s vector has already listened to its neighbors before the average is taken, which is why “not” moves the room more than a single voice would, and the smaller model’s curve flattens rather than vanishing. If models that read the whole passage at full attention become cheap enough to run on every document, the room stops mattering and this piece describes a stage, not a law.

09Back at the table

Your friend said “can’t.” It was one short word in a loud room, surrounded by the words you were listening for: Saturday, the boat, the lake. You heard the room.

He was only saying one thing, and it was the small word.

Background

• H. P. Luhn, “The Automatic Creation of Literature Abstracts,” IBM Journal of Research and Development, 1958.
• C. J. van Rijsbergen, Information Retrieval, 2nd ed., ch. 2, 1979.
• Christopher Fox, “A stop list for general text,” ACM SIGIR Forum, 1989.
• Dan Jurafsky and James H. Martin, Speech and Language Processing, 3rd ed. draft (open and closed word classes).
• Allyson Ettinger et al., “Assessing Composition in Sentence Vector Representations,” COLING 2018.
• Tristan Thrush et al., “Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality,” CVPR 2022.
• Mert Yuksekgonul et al., “When and why vision-language models behave like bags-of-words, and what to do about it?” ICLR 2023.
• Kyle Richardson et al., “Probing Natural Language Inference Models through Semantic Fragments,” AAAI 2020.
• Orion Weller, Dawn Lawrie and Benjamin Van Durme, “NevIR: Negation in Neural Information Retrieval,” EACL 2024.
• O’Connor v. Oakhurst Dairy, 851 F.3d 69 (1st Cir. 2017); settlement report, NPR, 2018.
• Models measured: all-MiniLM-L6-v2 and all-mpnet-base-v2 (both mean-pooled), sentence-transformers.
• John Rector, “Not Is a Nudge,” September 2026.
• John Rector, “Meaning Becomes an Address,” September 2026.
• John Rector, “Every Trap Is Near the Floor,” September 2026.

More at johnrector.me.