AI vocabulary · the whole list
The overused AI words list, sorted by evidence
Type this query into any search engine and the results are the same list of thirty to fifty words, copied forward from list to list with no source attached. This page is built from our own evidence instead: the checklist our linter actually runs, the model logs we measured ourselves, and the social-sentiment research that told us which popular claims hold up and which don't.
underslop splits overused AI vocabulary into three evidence tiers instead of one flat list. Four words, tapestry, testament, groundbreaking, and game-changer, are scored outright. A wider set of nineteen words, leverage, foster, robust, and others, is computed and reported but never counted, because they're ordinary technical English used sparingly. A third set, the classic essay-mill connectives like furthermore and moreover, turned up zero times across our own twelve-model conversation logs, which means most 'words to avoid' lists are still built around a fingerprint that has already expired.
What reaching for it signals
Across all three tiers, the underlying move is the same: reaching for a word instead of doing the specifying work the sentence actually needs. A scored word like tapestry substitutes texture for an unnamed connection. A greylist word like foster or leverage substitutes a soft causal verb for a described mechanism. The classic connectives, furthermore, moreover, in conclusion, substitute transition-sounding filler for an actual logical link between two sentences. None of these are wrong in isolation. What marks a paragraph as thin is stacking several of them in place of specifics, which is exactly the pattern a linter can catch and a human reader can't always name in the moment.
Whether underslop counts it
Not every word on a typical 'AI words' list deserves the same weight, and treating them as equally damning is the mistake this page exists to correct. Four words get scored because the evidence for them is layered: independent confirmation across our legacy skills plus social-sentiment data plus their own structural mechanism. Nineteen more get computed and reported, never scored, because they're ordinary technical vocabulary that reads fine anchored to a real referent. A further set of classic markers, the ones most other lists lead with, we searched for directly in our own twelve-model corpus and found essentially nowhere. Two of our four scored words are in that same zero-hit set. We score them anyway: reader complaints and legacy-skill evidence are still valid signals even when today's models have moved on, and a word can be a stale example of the problem without being a false one.
The list, split by how good the evidence is
Most published lists of AI words are one undifferentiated pile. Ours is tiered by what we actually measured across twelve models, because the tiers behave differently in real writing.
- Scored: counted against the run4 words
tapestry, testament, groundbreaking, game-changer
These four are the only vocabulary items underslop's linter subtracts points for outright. Each is confirmed across at least two of our legacy skills and independently named in our social-sentiment research as the most-cited complaint in its category. Mechanism write-ups live at /words/tapestry and /words/delve, which explains why delve is not on this list.
- Computed, never counted: the greylist19 words
robust, leverage, seamless, elevate, streamline, crucial, comprehensive, harness, foster, empower, innovative, journey, navigate, unlock, holistic, myriad, plethora, nuanced, synergy
Every run reports these for visibility, but none of them lower the score. They're ordinary technical or business English that reads fine when it's anchored to a specific referent and reads thin when it's stacked as filler. Foster gets its own page for the mechanism behind why it lands here rather than on the scored list.
- Found essentially absent in our own logs7 words
furthermore, moreover, additionally, in conclusion, it is important to note, stands as a testament, boasts
We searched our full twelve-model, roughly thirty-conversation corpus for this classic essay-mill cluster and came back at zero hits. The same search also came back at zero for delve, tapestry, robust, and foster as a raw frequency count. Most published 'AI words' lists still lead with this exact set. The words haven't disappeared from public complaint, they've disappeared from what current models actually write, which is a different and more interesting finding than the lists let on.
What to write instead
- a Tier 1 scored word (tapestry, testament, groundbreaking, game-changer)name the actual connection or result the word was standing in for
- These four cost a point on the linter outright. No context makes them free.
- several greylist words stacked in one paragraph (leverage, foster, unlock, elevate)the plain verb that names what actually happened
- None of these are individually scored, but a paragraph leaning on four of them in a row still reads thin, even when the linter says nothing.
- deleting delve because a list told you tokeep it if it's your natural register
- delve is reported, never scored, precisely because the popular case against it doesn't hold up. Full argument at /words/delve.
- copying a blocklist wholesale from another sitecheck it against current model output before you trust it
- Several of the classic entries on most lists turned up zero times in our own twelve-model search.
Examples
Our platform leverages a robust, comprehensive tapestry of tools to foster innovation and empower teams, standing as a testament to what a truly groundbreaking, game-changer approach can unlock in today's fast-paced, ever-evolving landscape.
Our platform gives teams a full set of tools to work with, and it was built to help them innovate. That is the straightforward version of what we offer.
This groundbreaking release will foster deeper engagement and empower our community to unlock a truly seamless, robust experience as we navigate an ever-evolving landscape together.
This release gives the community a smoother, more reliable experience. That should mean people use it more, which is the point.
Questions
What words are overused by ChatGPT?
Depends on the evidence you want behind the claim. underslop scores four outright (tapestry, testament, groundbreaking, game-changer), computes and reports nineteen more without scoring them, and tracks a third set that turned out to be nearly absent from our own model logs despite showing up on every other list.
What words give away ChatGPT?
Vocabulary alone rarely does. Our own research found structural habits, opening every reply by validating the user, converting casual conversation into bullet lists, closing with a cheerleading well-wish, present in every one of twelve models we sampled. Single words are easier to spot and easier to write a listicle about, which is most of why they get the attention.
Is leveraging a ChatGPT word?
It's on our greylist: computed and reported, never scored. Leverage is ordinary business vocabulary that reads fine anchored to something concrete, like 'leverage the existing customer list', and reads thin as unanchored filler, like 'leverage synergies'. The second use is the actual problem, not the word.
Is multifaceted a ChatGPT word?
It's not on our scored list or our nineteen-word greylist. It shows up often enough in social complaints that it's worth avoiding by ear, but we don't have the layered evidence for it that we have for the four words we actually score.
What are the top 10 overused words?
There isn't a single ranked ten on this site, because ranking implies one list is right. The honest version is three tiers by evidence strength, laid out above. The biggest surprise in our own data was not a new word worth adding to the list. Our own twelve-model search kept turning up zero hits for entries every other list still leads with.
underslop scores a small set of vocabulary tells and computes a larger greylist it deliberately does not count. Paste a draft to see which of the two your writing is tripping. The free tier caps at 500 words a run.