AI Slop Vocabulary — The Cut List

The phrases below trigger AI-detection radar instantly. Each one is documented in published research, AI-detector vendor methodology, or community-sourced lists with high agreement. The mechanical scanner (scripts/scan.py) catches every instance; this file is the human-readable reference with replacements and rationale.

Use this as a search-and-destroy list during a final pass. For each hit, ask: does this word survive the context? In almost every case, the answer is no — cut it.

How severity works

• H (always cut): the phrase is essentially never the right choice. Cut without exception unless ironic / scare-quoted.
• M (usually cut): the phrase survives in narrow bands. Default is to cut; keep only if doing specific work.
• L (context-dependent): weak tell on its own. Note in audit but don't down-score.

Density matters more than individual instances. See calibration.md.

---

2A. LLM-favored verbs

These are the verbs models reach for first. They sound active and serious without committing to a specific action. Most can be swapped for a one-syllable Anglo-Saxon verb that does the same work with less smell.

Word/Phrase	Why it's a tell	Replacement	Severity
delve / delve into	"Delves" appeared 6,697%+ more in 2024 PubMed vs 2020 — the flagship AI verb	look at, get into, study	H
leverage	Corporate cliche; appears in every blacklist	use	H
harness	Rare in human prose, ubiquitous in AI	use, channel	H
foster	Corporate-NGO speak	encourage, build	H
empower	Marketing fluff	help, give X to	H
unlock	"Unlock the potential of" is a top-10 GPT phrase	reveal, find	H
elevate	Empty intensifier	raise, lift, improve	H
streamline	McKinsey-speak	simplify, cut	H
revolutionize	Hyperbole baseline	change	H
transform	Same	change, rebuild	H
underscore / underscores	904% spike post-ChatGPT (PubMed study)	show, prove	H
illuminate	Decorative academic	clarify, show	H
navigate	"Navigate the complexities of" — top-10 cliche	handle, manage, work through	H
garner	Spike-word in academic writing studies	get, win, attract	H
utilize	Sounds smart, means "use"	use	H
facilitate	Same	help	H
optimize	Tech-jargon default	improve, tune	H
enhance	Generic intensifier verb	improve, sharpen	H
embark / embark on	"Embark on a journey" is iconic AI	start	H
showcase / showcasing	9.2x more frequent in AI than human (GPTZero)	show, display	H
boast / boasts	Wikipedia-flagged copula avoidance	has	H
demystify	LinkedIn cliche	explain	M
ignite	Copy-cliche	start, spark	M
supercharge	Marketing-speak	speed up	M
unleash	Same	release	M
unveil	Press-release verb	reveal, show	M
explore	Used to fill space ("we'll explore")	look at, go through	M
dive into	"Let's dive into" — sycophant cluster	start, look at	H
resonate / resonates	Vague impact-word	match, connect	M
reverberate	Even more decorative	echo	M
transcend / transcends	Often "transcends mere X"	go beyond	M
spearhead	Press-release cliche	lead	M
reimagine	Tech-deck cliche	redesign	M
craft	Overused as a verb	make, write	L
pave the way	Cliche metaphor	enable, set up	H
shed light on	Cliche	explain, show	H

Audit instruction: any H-tier verb is a hit. For M-tier, replace and re-read; if the sentence is unchanged or sharper, the verb was AI smell. Cut.

---

2B. Cliché metaphors and grandiose nouns

These nouns convert a small subject into an epic one. AI defaults to them because metaphor scores well in training data; humans use them sparingly because they sound like a press release.

Word/Phrase	Why it's a tell	Replacement	Severity
tapestry	Iconic AI metaphor	mix, weave (sparingly), or specific noun	H
landscape	"The landscape of X" — top cliche	field, market, world	H
realm	Decorative for "area"	area, field	H
beacon	"A beacon of X"	example, leader	H
treasure trove	Always cut	collection, source	H
symphony	Pretentious metaphor	combination, mix	H
journey	Especially "embark on a journey"	path, process	H
roadmap	Tech-deck cliche	plan, steps	M
ecosystem	Tech-cliche for "industry"	industry, network	M
paradigm / paradigm shift	Overused	change, shift	H
testament	"A testament to"	proof, evidence	H
cornerstone	Cliche	basis, anchor	M
crucible	Decorative	test, trial	M
labyrinth	Decorative	maze, complexity	M
metropolis	Travel-guide cliche	city	M
enigma	Decorative	mystery, puzzle	M
myriad / a myriad of	Inflated "many"	many, lots of	H
plethora	Same	lots of, too many	H
kaleidoscope	Decorative metaphor	mix, range	M
arena	Cliche metaphor	field	M
arsenal	Cliche metaphor	toolkit	M

Audit instruction: if you find one of these, ask whether the sentence needs a metaphor at all. Most don't. When the answer is yes, pick a domain-specific image — the model defaulted to the most-trained metaphor; you can do better.

---

2C. Empty intensifiers / hedges / vague adjectives

The biggest category by far. These adjectives are all reach and no grip — they assert importance without earning it. Density is what makes them lethal: one "crucial" survives; three in a paragraph guarantees AI.

Word/Phrase	Why it's a tell	Replacement	Severity
crucial	Top-3 spike word	important, key, or cut	H
essential	Same	needed, central	H
vital	Same	needed	H
pivotal	Same	key, decisive	H
paramount	Even more inflated	most important	H
profound	"A profound impact"	big, deep	M
robust	Tech-cliche	strong, reliable	H
seamless	Tech-cliche	smooth	H
comprehensive	Used to inflate completeness	full, complete	H
holistic	Buzz-word	whole, full	M
multifaceted	Always inflated	complex, many-sided	H
nuanced	Used as a flex	complex, subtle	M
intricate / intricacies	611% spike post-ChatGPT	complex, detail	H
meticulous / meticulously	Top spike word	careful	H
compelling	Marketing-cliche	strong, persuasive	M
commendable	Pretentious	good, deserving praise	M
insightful	Often empty praise	useful, sharp	M
invaluable	Hyperbole	useful, important	M
unwavering	Always inflated	steady, firm	H
transformative	Always inflated	major	H
groundbreaking	Always inflated	new, original	H
cutting-edge	Cliche	new, current	H
state-of-the-art	Same	new, top	H
game-changer / game-changing	Cliche	big change	H
next-generation	Tech-cliche	new	M
future-proof	Tech-cliche	lasting	M
dynamic	Vague intensifier	active, fast-changing	M
vibrant	Travel-guide cliche	lively, busy	M
bustling	Same	busy	M
daunting	Cliche	hard, intimidating	M
ever-evolving / ever-changing	"In the ever-evolving landscape"	changing, shifting	H
ever-expanding	Same family	growing	M
timeless	Inflated cliche	lasting	M
enduring	Often empty	lasting	M
diverse / diverse array of	Empty filler	mixed, varied	M
unique blend	Marketing cliche	mix	M
fast-paced	"In today's fast-paced..." 107x AI	fast, busy	H
hyper-connected	Cliche	connected	M
modern / today's	Often filler	now, current	L

Audit instruction: for each instance, try deleting it. If the sentence is fine or stronger without the word, it was filler. If a replacement is needed, prefer the shortest, most concrete word in the column.

---

2D. Sycophantic openers / closers

Direct RLHF artifacts. Every one of these is a model performing helpfulness instead of being helpful. Even when they leak into prose meant for publication, they read as machine-trained politeness.

Phrase	Why it's a tell	Replacement	Severity
Great question!	RLHF flattery	delete	H
Excellent question!	Same	delete	H
I'd be happy to help	Same	delete and answer	H
Absolutely!	Opener flattery	delete	H
Certainly!	Same	delete	H
Of course!	Same	delete	H
Sure! Here's...	Same	delete the opener, keep "Here's" if needed	H
I hope this helps!	Closing flattery	delete	H
Let me know if you have any questions	Same	delete	H
Feel free to reach out	Same	delete	H
Don't hesitate to ask	Same	delete	H
Is there anything else I can help you with?	Same	delete	H
I hope this answers your question	Same	delete	H
Happy to clarify	Same	delete	H
Let me know if you'd like me to elaborate	Same	delete	H

Audit instruction: every instance is a high-severity violation. Cut without exception. End on the last load-bearing sentence; openers go in the trash.

---

2E. Vague-authority phrases

Wikipedia's number-one content-pattern flag. These phrases assert evidence without citing any. AI defaults to them when it lacks specifics; humans either name a source or admit they're guessing.

Phrase	Why it's a tell	Replacement	Severity
Studies show	Uncited authority claim	name the study or cut	H
Research suggests	Same	name the research or cut	H
Many experts agree	Same	name the experts or cut	H
Industry reports indicate	Same	name the report or cut	H
It is widely understood	Wikipedia-flagged weasel	cut or attribute	H
Observers have noted	Anonymous authority	name the observer or cut	H
Some critics argue	Same	name the critic or cut	H
Generally speaking	Hedge filler	cut	M
In many cases	Hedge filler	cut or specify	M
It is commonly known	Weasel + filler	cut	M

Audit instruction: for each hit, either supply the citation (link, name, source) or cut the claim. "I think" beats "experts say" every time.

---

2F. Closing / connector clichés

The signposting words. AI was trained on five-paragraph essays and op-eds; it defaults to scaffolding even in short pieces. Real prose flows; it doesn't announce its turns.

Phrase	Why it's a tell	Replacement	Severity
In conclusion	Compulsive summary	cut	H
To conclude	Same	cut	H
In summary	Same	cut	H
To summarize	Same	cut	H
Overall	Same	cut	H
Ultimately	Filler closer	cut	M
All things considered	Same	cut	M
At the end of the day	Filler	cut	H
In essence	Restatement	cut	H
To put it simply	Restatement	cut	H
In a nutshell	Same	cut	M
Furthermore	Stock connector	period or "Also"	H
Moreover	Same	period or "Also"	H
Additionally	Same	period or "Also"	H
First and foremost	Listicle cliche	"First" or cut	H
Last but not least	Listicle cliche	"Finally" or cut	H
On the other hand	Stock contrast	"But"	M
That being said	Stock contrast	"Still" or cut	M
With that in mind	Filler transition	cut	M
Notably	Throat-clearing	cut	M
Indeed	Filler	cut	M

Audit instruction: most of these are deletable. End sentences with periods, not signposts. If a transition is genuinely needed, "But" / "Also" / "Still" carry their weight without smelling AI.

---

2G. Academically-validated spike words

These are the highest-confidence subset in the entire vocabulary list. Each one is statistically validated against pre-2022 baselines via published research — the spike is direct evidence that LLMs caused the surge in usage.

Word/Phrase	Spike data	Source
delves	6,697% increase 2020 to 2024	arXiv
underscores	904% increase	arXiv
intricate	611% increase	arXiv
showcasing	r=9.2 ratio AI vs human	arXiv
meticulous, meticulously	spike-confirmed	PubMed
pivotal	spike-confirmed	PubMed
commendable	spike-confirmed	PubMed
garnered	top-21 focal word	arXiv 2412.11385
boasts	top-21 focal word	arXiv 2412.11385
groundbreaking	top-21 focal word	arXiv 2412.11385
advancements	top-21 focal word	arXiv 2412.11385
aligns / aligns with	16x more frequent in AI	GPTZero
surpassing	12x more frequent in AI	GPTZero
impacting	11x more frequent in AI	GPTZero
play a significant role in shaping	182x more frequent in AI	GPTZero
today's fast-paced world	107x more frequent in AI	GPTZero
notable works include	120x more frequent in AI	GPTZero
aims to explore	50x more frequent in AI	GPTZero
objective study aimed	269x more frequent in AI	Atlas
research needed to understand	235x more frequent in AI	Atlas

A 2024 PubMed study estimated at least 13.5% of 2024 biomedical abstracts were processed with LLMs based on this excess vocabulary. After viral attention in early 2024, "delve" frequency in arXiv abstracts dropped sharply — confirming the words function as a fingerprint authors can sand off.

Audit instruction: any hit in this category is a near-certain AI marker. The data is statistical, not stylistic — these aren't bad words because they sound bad, they're bad because their frequency proves recent LLM authorship. Cut without exception.

---

How to use this list

1. Run the scanner first — scripts/scan.py flags every instance mechanically across all categories.
2. Walk the always-cut items (Severity H) — each hit is a high-severity violation. Cut without exception.
3. Walk the often-cut items (Severity M) — apply judgment. Default is to cut; keep only if doing specific work.
4. Note the context-dependent items (Severity L) — these flag stylistic register but don't down-score the verdict.
5. The list isn't exhaustive. New LLM tells emerge over time. If a phrase reads as engineered, it probably is. The default action is to cut.

The 2G category (academically-validated spike words) is the highest-confidence subset — these are statistically validated against pre-2022 baselines via published research.