Language models do not pick words evenly. They over-select a small set, and the effect is large enough that researchers can measure it against pre-2022 baselines and tell you the multiplier.
Below are thirty of the worst offenders, each with a replacement that carries more information than the word it replaces. The multipliers come from published corpus work, cited at the bottom.
Then the part most word lists leave out: why swapping these words will not, on its own, fix your writing.
The verbs
Verbs are where models do the most damage, because a vague verb hides the absence of a mechanism.
| Word | Measured | Use instead |
|---|---|---|
| delve / delves into | 28x expected rate (Kobak); +6,697% (Juzek) | examine, dig into, or just say what you found |
| underscore / underscores | 13.8x (Kobak); +904% (Juzek) | show, prove, highlight |
| showcase / showcasing | 10.7x (Kobak); +1,396% (Juzek) | show, demonstrate, display |
| leverage / leveraging | corporate register marker | use, apply, exploit |
| harness | same family as leverage | use, channel, run on |
| foster / fostering | vague non-action verb | build, encourage, cause, support |
| unlock / unleash | infomercial register | enable, start, allow, release |
| elevate | status inflation | raise, improve, lift |
| empower | hollow in most business use | let, allow, equip, pay for |
| streamline | process cliché | simplify, cut steps from, speed up |
| facilitate | almost always deletable | help, run, host, arrange |
| navigate | metaphor doing no work | handle, cross, deal with, get through |
| embark on | melodramatic | start, begin |
| shed light on | worn metaphor | explain, reveal, show |
How to use this table. Read the replacement column and notice what changes. "Facilitate a meeting" becomes "run a meeting", and now you know who was in charge. "Leverage our platform" becomes "use our platform", and the sentence has lost nothing except a suggestion of sophistication that was doing no work.
The modifiers
Adjectives are where slop hides its lack of evidence. Each of these asserts a quality without measuring it.
| Word | Why it flags | Use instead |
|---|---|---|
| robust | rarely means anything outside statistics | reliable, or state the failure rate |
| seamless / seamlessly | almost always untrue | quick, automatic, or name the step removed |
| multifaceted | says "complicated" at four syllables | complex, or list the facets |
| comprehensive | claims completeness without proof | complete, or say what is covered |
| cutting-edge / state-of-the-art | dates badly, proves nothing | new, or name the version |
| transformative / groundbreaking | superlative without evidence | say what changed and by how much |
| pivotal / crucial | asserts importance | important, or explain the consequence |
| meticulous | self-congratulation | careful, or describe the check performed |
| invaluable / unparalleled | unfalsifiable | useful, or give the number |
| ever-evolving | filler modifier | changing, or say what changed |
| vibrant / dynamic | atmosphere words | busy, fast, growing |
| nuanced | often means "I did not explain it" | complicated, or state the distinction |
The metaphor props
These are the ones readers mock, because they arrive in contexts where no metaphor was needed.
| Word | Measured | Use instead |
|---|---|---|
| tapestry / rich tapestry | "vibrant tapestry" at ~17,000x baseline (Pangram) | mix, range, or just list the things |
| testament / stands as a testament | "serves as a testament" ~4,000x (Pangram) | shows, proves |
| beacon | faux-reverent | example, model, guide |
| landscape | metaphorical crutch | market, field, industry |
| realm | pompous framing | field, area |
What a fixed paragraph looks like
Take a real sentence carrying six of these:
Our robust platform leverages cutting-edge technology to seamlessly empower teams to navigate the ever-evolving digital landscape.
Word-swapping alone gets you here:
Our reliable platform uses new technology to quickly let teams handle the changing digital market.
Grammatically fine. Still says nothing, because the problem was never the vocabulary. Every noun in that sentence is unspecified.
The actual fix requires knowing what the thing does:
Our tool reads your Jira board and assigns pull request reviewers in under ten seconds. In a sixty-day trial with forty companies, review time fell 35 per cent.
That is the difference between a word list and an editing method. A list tells you which words to suspect, which is useful for about ten minutes and then stops being the bottleneck, because the sentence above failed on evidence rather than vocabulary. No list can tell you what your product does.
Why swapping words is not enough
Three reasons the list on its own will disappoint you.
Rhythm survives the swap. Replacing "delve" with "examine" leaves your sentence the same length. Machine writing clusters between 14 and 18 words per sentence and stays there, which readers register as monotony before they notice a single word. Edited human prose scores above 0.60 on sentence-length variance. Raw model output sits near 0.30. No amount of vocabulary editing moves that number.
The structures survive too. The invented third item in a list. The "not just X, it's Y" construction. The closing paragraph that summarises what you just read. None of those are words, so none of them appear on a word list.
Density is untouched. A paragraph with zero checkable claims still has zero after you improve its vocabulary. This is the deepest tell and the only one that cannot be faked, because it requires you to know something.
Doing it before the draft instead of after
Editing these out afterwards costs about 45 minutes per piece. The cheaper order is to stop the model producing them.
Models follow constraints of a specific shape. Give one an explicit list of banned tokens and it complies. Tell it to vary sentence length deliberately, cap em dashes at one per 500 words, and require each section to carry a claim a reader could check, and it complies with those too.
Ask it to "write more naturally" and nothing happens, because natural is not something a model can act on at the token level. It is a judgment about output, not an instruction about generation.
The thirty words above are a working subset. Our full lexicon runs to 311 terms with replacements for each, and the free workbench runs about half of them against your own writing, flags what it finds in context, measures your sentence variance against the human threshold, and returns a score. Paste in a post you have already published. The result is usually informative.