AI words list: the 100 words and phrases AI overuses (measured)

We asked three current AI models the same 300 real questions that humans had already answered, then counted every word and phrase. This is what AI writing over-uses in 2026, with numbers. Updated 2026-11-12.

Downloads (CC BY 4.0): CSV · JSON · Methodology

Key findings

Em dashes are the biggest tell. 4.5 per 1,000 words in AI answers vs 0.21 in human answers (~21×). gpt-6-luna hit 8.06. ChatGPT in 2022 used 0.06: this habit is new. More on em dashes.
Formatting gives AI away before the words do. 29.24 bullet lines and 32.68 bold spans per 1,000 words, against 0.03 and 0 for humans.
The new verbal tic is the offer. "If you tell me…" appeared 34.1× more in AI answers and in 11.3% of them; humans almost never write it. "Here are some", "I can explain" and "if you want, I can" follow the same pattern.
Hedging is everywhere. "may" (4.3×), "often", "typically", "sometimes", "especially if", "can lead to".
"Delve" is dead. Zero uses of delve or tapestry in 67,773 words of modern AI text. Lists of "AI words" built in 2023 are out of date.

Check your own text

0 words · try a sample

Free: 3 AI rewrites a day (up to 300 words) and unlimited quick clean. Need more? A $5 pack of 50 rewrites never expires, or go Pro at $12/month · $79/year. See plans.

Top 100 words and phrases AI overuses

Ratio = AI rate ÷ human rate (add-one smoothed). Rates are occurrences per 10,000 words. "% AI answers" = share of the 300 AI answers containing it.

#	Word / phrase	AI vs human	AI /10k	Human /10k	ChatGPT 2022 /10k	% AI answers
1	tell me	26.8×	7.97	0.15	0	18%
2	if you tell me	34.1×	5.02	0	0	11.3%
3	and whether	28.2×	4.13	0	0.29	9%
4	can help	13.6×	8.12	0.45	8.37	12.7%
5	can lead to	17.5×	5.16	0.15	3.61	7.3%
6	specific	8.2×	12.25	1.36	9.96	21.3%
7	summary	19.5×	2.8	0	0.72	6%
8	here are some	12.2×	3.54	0.15	1.3	8%
9	common	6.7×	14.02	1.97	4.47	21.7%
10	whether	6.5×	13.72	1.97	2.89	21.7%
11	it's important to	16.6×	2.36	0	9.09	5%
12	plan	9.5×	5.61	0.45	4.47	9%
13	can sometimes	16.6×	2.36	0	0.43	4.7%
14	avoid	8.4×	7.53	0.76	4.62	10.7%
15	may	4.3×	45	10.31	54.55	48.7%
16	and may	15.6×	2.21	0	3.61	4.7%
17	rules	6.7×	11.07	1.52	2.6	13.7%
18	improve	10.2×	2.95	0.15	6.78	6.3%
19	here are	6.3×	5.61	0.76	3.17	12.3%
20	especially if	9.7×	2.8	0.15	1.3	6%
21	beneficial	13.6×	1.92	0	0.58	4%
22	lead to	7×	7.23	0.91	5.2	9%
23	important to	7.8×	3.39	0.3	26.55	7.3%
24	if you want i	12.7×	1.77	0	0	4%
25	can create	12.7×	1.77	0	0.87	4%
26	practical	9.2×	2.66	0.15	0.43	5.7%
27	factors	6.5×	5.75	0.76	7.79	9.7%
28	often	3.5×	30.99	8.79	17.32	42.3%
29	strongly	13.6×	1.92	0	0	3.3%
30	crucial	11.7×	1.62	0	0.14	3.7%
31	especially	3.9×	12.25	3.03	2.02	23.3%
32	conditions	5.4×	6.34	1.06	6.35	10.3%
33	help	3.8×	17.71	4.55	22.22	24%
34	typically	4.2×	8.71	1.97	9.24	17.3%
35	expectations	12.7×	1.77	0	0.29	3%
36	changes	4.3×	9	1.97	5.05	14.7%
37	i can explain	6.5×	2.8	0.3	0	6.3%
38	and avoid	10.7×	1.48	0	0.72	3.3%
39	sometimes	3.5×	15.94	4.4	1.73	23.3%
40	contribute	7.8×	2.21	0.15	1.73	4.3%
41	vary	5.1×	3.69	0.61	2.45	8.3%
42	can cause	6.3×	3.69	0.45	5.2	5.3%
43	steps	5.8×	3.39	0.45	1.88	6%
44	check the	7.3×	2.07	0.15	0	4.3%
45	you mean	5.6×	3.25	0.45	0.14	6.3%
46	timing	6.8×	2.95	0.3	0.14	4.7%
47	can change	9.7×	1.33	0	0.43	3%
48	vary by	9.7×	1.33	0	0.14	3%
49	safe	4.6×	5.46	1.06	2.89	8.7%
50	requirements	7.3×	2.07	0.15	2.02	4%
51	whether it	6.8×	1.92	0.15	0.43	4.3%
52	weight	7.8×	2.21	0.15	3.17	3.7%
53	anxiety	8.3×	2.36	0.15	0.87	3.3%
54	consider	4×	8.26	1.97	9.52	11%
55	typical	6.2×	2.66	0.3	0.14	4.7%
56	situation	4.2×	5.61	1.21	4.18	9.3%
57	risks	7.3×	2.07	0.15	2.74	3.7%
58	contribute to	9.7×	1.33	0	1.59	2.7%
59	exercise	8.3×	2.36	0.15	0.72	3%
60	spread	5.5×	2.36	0.3	1.01	5%
61	key	3.3×	7.38	2.12	2.6	15.7%
62	patterns	6.8×	1.92	0.15	2.31	3.7%
63	ensure	4.5×	4.57	0.91	2.89	7%
64	create	3.7×	5.46	1.36	6.35	10.7%
65	monitoring	8.8×	1.18	0	0.14	2.7%
66	tailored	8.8×	1.18	0	0.29	2.7%
67	example if	8.8×	1.18	0	2.02	2.7%
68	etc i can	8.8×	1.18	0	0	2.7%
69	professional	4.4×	3.84	0.76	5.63	7%
70	issues	3.9×	6.34	1.52	3.61	9%
71	offer	4.2×	3.69	0.76	3.17	7.3%
72	routine	6.8×	1.92	0.15	0.14	3.3%
73	include	3.3×	6.93	1.97	9.38	13%
74	check	3.3×	7.38	2.12	2.02	13%
75	injury	7.3×	2.07	0.15	1.73	3%
76	or other	3.8×	3.84	0.91	6.64	8.7%
77	manage	4.3×	3.1	0.61	4.33	6%
78	understanding	4.3×	3.1	0.61	2.02	5.7%
79	reduce	3.4×	6.05	1.67	5.92	9.3%
80	pattern	5.8×	1.62	0.15	0.29	3.3%
81	mainly	4.9×	2.07	0.3	0	4.3%
82	commonly	3.5×	3.54	0.91	2.74	7.7%
83	prices	3.9×	6.34	1.52	3.17	5.7%
84	health	3.4×	6.05	1.67	6.64	7.7%
85	period	4.1×	3.54	0.76	4.47	5%
86	quick	3.3×	3.39	0.91	0.43	7.7%
87	stress	3.9×	3.98	0.91	1.15	5.3%
88	underlying	4.5×	1.92	0.3	3.17	4%
89	provider	5.1×	2.95	0.45	6.49	3.3%
90	safety	3.9×	3.98	0.91	1.15	5%
91	allowing	3.9×	2.8	0.61	0.72	5%
92	long term	4.6×	2.66	0.45	2.31	3.7%
93	affect	3.3×	3.39	0.91	2.89	7%
94	healthcare	3.5×	4.13	1.06	9.81	6%
95	goals	4.9×	2.07	0.3	3.9	3.3%
96	compare	4.5×	1.92	0.3	1.3	3.7%
97	near	3.6×	3.1	0.76	0.14	5.7%
98	details	3.9×	2.21	0.45	0.72	4.7%
99	describe	4.2×	1.77	0.3	0.43	4%
100	require	3×	4.87	1.52	1.73	8.3%

Formatting habits

Per 1,000 words	Modern AI (all)	Human	ChatGPT (2022)	gpt-6-luna	gpt-5.4-nano	gpt-4o-mini
Em dashes (—)	4.5	0.21	0.06	8.06	5.9	0.5
Bullet / numbered lines	29.24	0.03	0	13.16	42.61	24.29
Bold spans (…)	32.68	0	0	23.61	47.58	21.28
Markdown headings	6.48	0	0	1.26	13.23	2.05
Colons	19.52	3.24	4.96	9.95	19.88	25.46
Semicolons	1.28	1.21	0.04	2.83	1.29	0.25
Exclamation marks	1.42	0.56	0.48	0.19	0.46	3.35
Mean sentence length (words)	20.1	19.8	25.7	17.6	23.2	18.8

The classic "AI slop" words, measured

The words everyone jokes about are rarer than you'd think in 2026 Q&A answers. Per 10,000 words:

Phrase	Modern AI	ChatGPT 2022	Human
overall,	4.28	8.51	1.36
in summary,	2.21	0.72	0
ultimately,	0.3	1.44	0.61
additionally,	0.3	1.15	0.15
testament	0.3	0.58	0
leverage	0.74	0	0.15
i hope this helps	0	0.72	0
vibrant	0.3	0.29	0
landscape	0.44	0.14	0
when it comes to	0	0.58	0.45
foster	0.15	0.43	0.15
robust	0.44	0	0.15
in conclusion,	0.3	0	0
importantly,	0.3	0	0.45
utilizes	0.3	0	0

They still show up in marketing copy, LinkedIn posts and essays, which is why the unslop rule engine keeps all of them. The full rule list (100+ phrases with plain replacements) is free: rules CSV · rules JSON.

Methodology

Questions: 300 questions sampled (fixed seed) from the HC3 dataset (CC BY-SA 4.0): 120 reddit eli5, 60 finance, 60 open qa, 30 medicine, 30 wiki csai.
Human text: the human answers HC3 collected for those questions (65,983 words), written before ChatGPT's launch.
AI text: generated on 2026-11-12 with no system prompt and default settings: 100 by gpt-4o-mini, 100 by gpt-5.4-nano, 100 by gpt-6-luna (67,773 words). We also report HC3's original ChatGPT (Dec 2022) answers as a third column (69,297 words).
Counting: lower-cased words, markdown symbols stripped, 1–4 word n-grams. Kept phrases that appear in at least 8 AI answers, across at least 3 of the 5 topic sources (no source over 60%), with an AI/human ratio of at least 3. Overlapping n-grams were merged; ranking = log(ratio) × log(answers containing it).
Limits: Q&A answers, not essays or marketing copy; some topic words (health, prices) remain because AI answers go into more detail than Reddit replies. Treat single words as signals, not proof. No detector can prove authorship.

FAQ

What words does AI use the most?

In our 300-question study the words and phrases most over-used by modern AI compared with humans answering the same questions were offers like "if you tell me" (34.1× the human rate), "it's important to", "can help", "can lead to", hedges like "may", "often" and "typically", and words like "specific", "common", "crucial", "beneficial" and "tailored".

Does ChatGPT still say delve?

Not much. In 67,773 words of 2026 AI answers we found zero uses of "delve" or "tapestry". The classic slop words have mostly been trained out; the new tells are formatting (bold, bullets, em dashes) and hedged, offer-style phrasing.

Do AI models really overuse em dashes?

Yes. Modern AI answers used 4.5 em dashes per 1,000 words vs 0.21 for humans (about 21×). ChatGPT in 2022 used almost none (0.06), so the habit is recent.

Can I reuse this data?

Yes, under CC BY 4.0: link to gounslop.com/ai-slop-phrases. The CSV and JSON downloads have every number on this page.