The Word Your AI Quietly Redefined
A single word in your prompt can mean one thing to you and another thing to the model. That gap costs you finished work. The fix is not a bigger glossary. It is a written definition paired with behavior that matches it.
A clear word that the model read differently
Here is the case that started this for me. I maintain a specification for an AI assistant. For months I would say “let’s do a follow-up task.” Every time, the assistant treated the follow-up as smaller than the original.
The main work would reach about 80 percent. The last 20 percent moved into the next task. That task also stalled at partial. Its leftovers moved again. The pattern repeated, and the work never reached 100 percent.
No human would misread “follow-up.” To me it means the next unit of work. Same weight. Sometimes larger. The model had quietly attached a second meaning: a small addendum. I never wrote that meaning down. It still governed the output. That is the trap. An unwritten definition is still a definition to the model.
New research shows words carry internal weight
On July 6, 2026, Anthropic’s interpretability team published “A Global Workspace in Language Models” (post, paper). It describes what they call the J-space. That is a small, privileged set of internal patterns where the model holds what it is currently thinking about.
Two numbers make it concrete. The workspace holds only a few dozen concepts at a time. It accounts for less than a tenth of the model’s internal activity. Each pattern is tied to one word. When a pattern lights up, the researchers write, it does not mean the model is saying that word, “just that the word is on its mind.”
So the instinct held up. A term you feed a model is not inert. It carries internal structure, and that structure has causal weight on the output.
This is not one vendor’s quirk either. An independent group replicated the core finding on an open Qwen model. The result was less clean, but it held. The tool they used runs on open decoders you can download today. The effect travels across models, which is what matters if you write for more than one.
You do not need the neuroscience
Here is where I step off the research. You do not need the J-space to justify the fix.
The mechanism is old and ordinary. What I built is a controlled vocabulary. Librarians have used the idea for decades. The W3C standard for it, SKOS, dates to 2009. The broader semantic-web work goes back to 2001. My contribution is not the concept.
The J-space is motivation, not proof. It made me take the friction seriously. It does not prove my design is correct. I want to say that plainly. Most write-ups that touch this paper slide into “the AI is conscious.” The paper is careful. It studies access consciousness, meaning what the model can report and reason with. It makes no claim about subjective experience. Neither do I.
There is even a caveat inside the caveat. A commentary on the paper argued the effective capacity may be closer to six coherent ideas than a few dozen. The hundred-fold connection-density figure appears in the blog, not the paper. So the practical lesson is smaller than the headline: keep your set of load-bearing terms small.
The fix is a definition plus matching behavior
Treat each load-bearing term the way a controlled vocabulary does. Five moves, in order.
Give it one canonical label. Pick a single word or phrase and stop paraphrasing it elsewhere.
Give it one owning place. Define it once, in the document that owns it, in plain text.
State what it is not. My “follow-up” entry now reads: “the next unit of work, equal weight by default; not a smaller addendum.” The negative half carries as much weight as the definition. It is the part the model had been getting wrong, so it is the part I wrote down first.
Put it to work. A definition that appears once and is never used again teaches the model nothing. Meaning lands through repeated use, not through a one-time declaration.
Then the move I missed for months. Make the process live the word. My bug was not a missing definition. The bug was that the behavior contradicted the word. The assistant said “follow-up” and then acted out “smaller.” A cleaner definition alone would not have fixed that. The rule that fixed it was behavioral: deferring the rest of the work is a decision the user makes, not a default the model takes.
That pairing is the whole trick. Definition plus congruent behavior. A glossary changes nothing if your process keeps doing the opposite of what the word says.
Audit your load-bearing terms this week
Open the prompt, spec, or agent instructions you rely on most. List the words that carry real weight. There are usually only a handful.
For each one, check three points. It has a single stated meaning. You have written down what it does not mean. Your own instructions never act against it.
You will probably find one word doing quiet damage, the way “follow-up” did for me. Fix the behavior, not only the wording. Then watch the next few runs for the term acting against its definition again. That costs you a paragraph and a little attention. It saves you the 20 percent that keeps going missing.