The delve index

Delve is down 98% from its arXiv peak, and a list of 30 AI-linked words is back near its pre-ChatGPT rate on arXiv, Hacker News and GitHub READMEs.

Cover for The delve index, styled as a dictionary. The word delve is struck through beside a huge em dash. Delve in arXiv abstracts is down 98% from its peak, 63.76 to 1.37 per million words, Dec 2023 to Feb 2024 against Aug to Oct 2026. The em dash is in 39% of sampled READMEs of GitHub repos created Jan to Oct 2026, against 1.9% in 2024.
On this page

For about a year, "delve" was the word people pointed at when they suspected a chatbot. They had a case. From December 2023 to February 2024 it appeared 595 times in 9.33M words of new arXiv abstracts: 63.76 per million, about 21 times its rate before ChatGPT. From August to October 2026 it appeared 29 times in 21.22M words. That's 1.37 per million, down 98% from the peak and 0.45× the 3.06 per million of January 2019 to October 2022.

Geng and Trotta already showed delve falling on arXiv in data through 2024. We wanted to know what happened to the rest of the vocabulary, so we wrote down a list of 31 words before counting anything, and ran it over arXiv abstracts, Hacker News comments and the READMEs of GitHub repos, through October 2026. We later took one word out, so the headline list is the other 30; here's why, with both versions. We looked for a cross-platform version of this and we didn't find one. The list is back near its pre-ChatGPT rate, but read that as suggestive: on arXiv ordinary control words moved as far, and on Hacker News the later fall may partly be moderation. These are word-frequency markers: they don't detect AI text, we don't estimate how much text AI wrote, and falling markers don't mean falling AI use. The methodology explains how we counted, and where the method falls short.

Key findings

  1. 1
    "Delve" on arXiv is down 98% from its peak. 63.76 per million words (Dec 2023 to Feb 2024) to 1.37 (Aug to Oct 2026), a ratio of 0.021 with a 95% interval of 0.012 to 0.031. That's 0.45× its pre-ChatGPT rate. The 98% is arXiv only.
  2. 2
    A list of 30 AI-linked words is back near its pre-ChatGPT rate on all three platforms, which is suggestive rather than proven. arXiv went 2,335 → 5,858 (2025) → 1,969, or 0.84×. READMEs went 871 → 2,770 (2024) → 643, or 0.74×. Hacker News, where the rise was small, went 594 → 673 (2024) → 601, or 1.01×. We registered 31 words and removed harness after seeing its result; with it in, Hacker News is 1.29×, which isn't near. On arXiv the return doesn't clear a strict placebo (finding 5). On Hacker News, dead comments went from 2.57% in 2025 to 5.67% from March 2026, so some of its fall may be removal.
  3. 3
    It happened in two phases. From 2024 to 2025, arXiv's delve fell 77% while the list kept rising: later words took over from earlier ones. In 2026 nearly every word on the list falls.
  4. 4
    34 of 40 registered markers on arXiv peaked and fell, against 8 of 30 control words. On Hacker News it's 15 of 40 against 1 of 30. Same rule, no selection.
  5. 5
    The arXiv controls moved too. Ordinary words ended anywhere from 0.32× to 3.2× their baseline in 2026, so the list's 0.84× doesn't clear a strict placebo. Delve's rise and fall do, and so does "genuinely".
  6. 6
    Exploratory: the em dash. In our sample of READMEs of GitHub repos created from January to October 2026, read as they stood 30 days after creation, 39.0% contain an em dash (34.6% across all repos, 44.5% among repos with at least one star). For 2024 repos it's 1.9%.
  7. 7
    Exploratory: "genuinely". Up 6.2× on arXiv and 1.9× on Hacker News, both clear of the controls, and from 3 uses to 29 in READMEs. It's the one registered word that isn't a topic word and is up on all three. It has started to slip.
  8. 8
    On Hacker News, delve isn't a few accounts, and in 2026 about half of it isn't the ordinary word. The top 10 authors hold 4 to 9% of uses every year. In a hand check, ordinary use went from 93.8% before ChatGPT to 52.5% in 2026 (63 of 120, interval 43.6% to 61.2%); the rest is names, quotes and people talking about the word, helped that year by one startup's news.

−98%

"Delve" in arXiv abstracts, peak (Dec 2023 to Feb 2024) to Aug to Oct 2026. arXiv only

Source: host0 analysis of arXiv abstracts

0.84×

The list of 30 on arXiv against its pre-ChatGPT rate, inside the range of the control words. Hacker News 1.01×, GitHub READMEs 0.74×

Source: host0 analysis

34 of 40

Registered markers on arXiv that peaked and fell, against 8 of 30 control words

Source: host0 analysis

39.0%

Sampled READMEs of GitHub repos created Jan to Oct 2026 with an em dash, at 30 days old. Exploratory

Source: host0 analysis

Delve on arXiv rose about 21×, then fell 98%#

We count every form of the word (delve, delves, delved, delving) in every field, and date each paper by its arXiv id. "Per million words", or ppm, is the unit throughout. The baseline is January 2019 to October 2022, which stops before ChatGPT's launch on 30 November 2022.

"Delve" in arXiv abstracts

Uses per million words, centred 3-month rate, all fields, Jan 2019 to Oct 2026

"Delve" in arXiv abstracts
xy
Feb 20190.91
Mar 20190.69
Apr 20191.14
May 20190.98
Jun 20191.13
Jul 20190.84
Aug 20191.3
Sep 20190.95
Oct 20190.92
Nov 20191.85
Dec 20192.96
Jan 20203.98
Feb 20203.06
Mar 20202.68
Apr 20202
May 20202.27
Jun 20202.32
Jul 20202.35
Aug 20203.42
Sep 20203.28
Oct 20203.58
Nov 20202.33
Dec 20202.53
Jan 20212.18
Feb 20214.06
Mar 20214.66
Apr 20214.5
May 20213.15
Jun 20212.9
Jul 20213.24
Aug 20212.9
Sep 20212.01
Oct 20212.59
Nov 20213.73
Dec 20214.48
Jan 20225.07
Feb 20225.01
Mar 20225.05
Apr 20224.07
May 20223.43
Jun 20223.38
Jul 20223.98
Aug 20225.55
Sep 20225.86
Oct 20226.08
Nov 20224.98
Dec 20225.64
Jan 20235.93
Feb 20237.25
Mar 20238.67
Apr 202311.38
May 202316.14
Jun 202318.92
Jul 202326.84
Aug 202334.98
Sep 202345.17
Oct 202352.52
Nov 202358.83
Dec 202363.61
Jan 202463.76 (63.76 · Dec 2023 to Feb 2024)
Feb 202462.94
Mar 202461.32
Apr 202458.47
May 202451.49
Jun 202448.62
Jul 202444.61
Aug 202441.18
Sep 202433.96
Oct 202429.96
Nov 202428.4
Dec 202425.37
Jan 202522.36
Feb 202517.4
Mar 202515.85
Apr 202513.86
May 202513.4
Jun 202510.79
Jul 20259.01
Aug 20256.06
Sep 20255.88
Oct 20255.67
Nov 20255.7
Dec 20254.87
Jan 20264.09
Feb 20263.8
Mar 20262.45
Apr 20262.03
May 20261.85
Jun 20262.09
Jul 20261.86
Aug 20261.17
Sep 20261.37
The peak window has 595 uses in 9.33M words. The last window, Aug to Oct 2026, has 29 uses in 21.22M words. The pre-ChatGPT rate was 3.06 (331 uses in 108.2M words). Each point is the three months around that month.

Source: host0 analysis of 1.75M arXiv abstracts

The count includes the DELVE astronomy survey, in capitals. Lowercase only tells the same story: 2.65 in the baseline, 44.6 in 2024, 0.99 in the latest window. A peak is the highest of many windows, so it's biased up a little; read the 98% with its interval, a ratio of 0.012 to 0.031.

It wasn't only computer science. At its peak, delve ran 15.6× its baseline in AI-type categories, 29× in other CS, 20.9× in physics, 23.7× in math, 28.7× in statistics and 34.8× in electrical engineering. Levels differ (108 ppm at the peak in other CS, 46 in physics, 37 in math), and the small fields have small counts, but the rate fell in all of them: the latest window is 0.5% to 5.4% of the peak where the counts allow a number.

Hacker News fell too, by a smaller amount. By our registered rule delve peaked and fell there, from 7.26 ppm (330 uses, centred on May 2024) to 2.61 (133 uses), a ratio of 0.36 with an interval of 0.29 to 0.44. That's 0.86× its baseline of 3.04. On Show HN (9 uses at the peak, 4 in the latest window) and GitHub READMEs (19 uses in 5.4M words, none in the latest window) the word is too rare to claim anything.

The whole list is back near where it started#

One word can fall for its own reasons, so the test that matters is a list. Ours comes from Wikipedia's advice page on signs of AI writing and the published literature, and we wrote it down before counting, except that one registered word came out later (see below): delve, tapestry, meticulous, intricate, pivotal, underscore, realm, leverage, robust, testament, garner, boast, fostering, showcasing, enhance, align with, emphasizing, highlighting, multifaceted, prioritize, crucial, landscape, vibrant, interplay, seamless, comprehensive, nuanced, navigate, genuinely and significant. The same 30 words and the same tokenizer run on every platform, and each platform is compared only with its own past. Levels between platforms reflect genre, so arXiv's 1,969 against Hacker News's 601 says nothing about who uses more AI.

The list of 30, against each platform's own pre-ChatGPT rate

Rate as a multiple of the Jan 2019 to Oct 2022 rate, by year, then the latest window

The list of 30, against each platform's own pre-ChatGPT rate
Seriesxy
arXiv abstracts20190.89×
arXiv abstracts20200.97×
arXiv abstracts20211.03×
arXiv abstracts20221.11×
arXiv abstracts20231.61×
arXiv abstracts20242.4×
arXiv abstracts20252.51×
arXiv abstractsSep 20260.84×
Hacker News comments20190.99×
Hacker News comments20201.02×
Hacker News comments20211×
Hacker News comments20220.99×
Hacker News comments20231.06×
Hacker News comments20241.13×
Hacker News comments20251.13×
Hacker News commentsSep 20261.01×
GitHub READMEs20190.98×
GitHub READMEs20201.04×
GitHub READMEs20211.1×
GitHub READMEs20220.9×
GitHub READMEs20232.01×
GitHub READMEs20243.18×
GitHub READMEs20253.08×
GitHub READMEsSep 20260.74×
1.00 is the pre-ChatGPT rate. The last point is Aug to Oct 2026 on arXiv and Hacker News, and Jul to Oct 2026 on GitHub. In ppm: arXiv 2,335 → 5,858 (2025) → 1,969; Hacker News 594 → 673 (2024) → 601; READMEs 871 → 2,770 (2024) → 643.

Source: host0 analysis of arXiv, Hacker News and a GitHub sample

We say "near" on purpose, because the answer depends on how you summarise the list. The median word, among those with at least 20 baseline uses (28, 30 and 16 words), sits at 0.88× on arXiv, 0.93× on Hacker News and 0.87× on READMEs. Drop the three most common words and it's 0.99×, 0.99× and 1.08×. Hold arXiv's field mix at its baseline shares and it's 0.85×; use only papers with one version, so the abstract is what was first submitted, and it's 0.86×. Physics alone is 1.10×. The honest summary is "back to roughly where it was". The 95% intervals on the latest rates are 1,945 to 1,993 on arXiv, 591 to 612 on Hacker News and 517 to 784 on READMEs.

READMEs are the shakiest of the three. July to September alone is 799 ppm (0.92×), and October, with under 30 days of history, is 469. READMEs also got longer, so the share with at least one listed word is up (14.9% to 23.8%), though among READMEs of 300 words or fewer it barely moves (10.6% to 11.4%). And on arXiv this return is suggestive, not proven, because ordinary words moved as far.

Hacker News deserves its own shape, because it never had the climb. Its comments moved 13% at most on a yearly view, and the fall since then may have a separate cause, which we get to later.

The list of 30 on Hacker News

Uses per million words in comments, monthly, Jan 2019 to Oct 2026

The list of 30 on Hacker News
xy
Jan 2019590
Feb 2019580
Mar 2019577
Apr 2019605
May 2019593
Jun 2019585
Jul 2019583
Aug 2019592
Sep 2019573
Oct 2019576
Nov 2019588
Dec 2019584
Jan 2020611
Feb 2020588
Mar 2020610
Apr 2020612
May 2020603
Jun 2020610
Jul 2020608
Aug 2020606
Sep 2020584
Oct 2020579
Nov 2020610
Dec 2020611
Jan 2021597
Feb 2021590
Mar 2021600
Apr 2021577
May 2021589
Jun 2021585
Jul 2021586
Aug 2021595
Sep 2021617
Oct 2021599
Nov 2021615
Dec 2021587
Jan 2022586
Feb 2022582
Mar 2022589
Apr 2022596
May 2022606
Jun 2022594
Jul 2022579
Aug 2022595
Sep 2022586
Oct 2022586
Nov 2022591
Dec 2022590
Jan 2023592
Feb 2023606
Mar 2023597
Apr 2023622
May 2023660
Jun 2023641
Jul 2023642
Aug 2023626
Sep 2023635
Oct 2023646
Nov 2023655
Dec 2023652
Jan 2024652
Feb 2024669
Mar 2024675
Apr 2024667
May 2024674
Jun 2024674
Jul 2024667
Aug 2024667
Sep 2024681
Oct 2024688
Nov 2024690 (690 · Nov 2024)
Dec 2024683
Jan 2025681
Feb 2025652
Mar 2025689
Apr 2025664
May 2025673
Jun 2025672
Jul 2025686
Aug 2025662
Sep 2025664
Oct 2025654
Nov 2025651
Dec 2025674
Jan 2026675
Feb 2026672
Mar 2026651
Apr 2026636
May 2026629
Jun 2026622
Jul 2026634
Aug 2026630
Sep 2026595
Oct 2026579
The axis starts at zero, so the small rise looks small. By period: 594 in the baseline, 673 in 2024, 601 in Aug to Oct 2026. October 2026 alone is 579.

Source: host0 analysis of 26.7M Hacker News comments

Show HN posts tell a sharper story but a less certain one. There the list fell 69% from 2024, from 1,779 ppm (1,687 to 1,874) to 555 (508 to 604). We can't compare it with the baseline: before about 2022 the dataset's Show HN records are mostly titles, so any index against those years is inflated.

Why 30 words and not 31#

Our registration lists 31 words in these groups. It doesn't define a list-level number at all; we added that summary afterwards, and we took one word out after seeing its result, so this is a departure from what we registered. The word is harness, and it turned out to be a topic: agent harnesses. On arXiv it went from 25 to 437 ppm in AI-type categories between January 2022 and October 2026, while staying flat in physics (14 to 19) and math (3 to 3). With it back in, the three platforms read 0.90×, 1.29× and 1.04×. The median word index is 0.88×, 0.93× and 0.87× without it, and 0.93×, 0.93× and 0.87× with it, so you can decide for yourself.

How much the answer depends on one word

Latest rate as a multiple of the pre-ChatGPT rate, three ways of summarising the list

How much the answer depends on one word
LabelValue
arXiv, list of 300.84×
arXiv, with harness0.90×
arXiv, median word0.88×
Hacker News, list of 301.01×
Hacker News, with harness1.29×
Hacker News, median word0.93×
READMEs, list of 300.74×
READMEs, with harness1.04×
READMEs, median word0.87×
Highlighted bars keep harness in; the median bars leave it out. With it, Hacker News sits at 1.29×, which isn't near its baseline. The median word barely notices, because one word can only move a median by one place. Harness stays in the 40-marker tally below, unchanged.

Source: host0 analysis

Two phases: a swap, then a fall#

Did delve fall while the other words rose? Yes, and then no. Harness aside, a list written down in advance is what lets us say so without picking the words afterwards.

2024 to 2025 was a swap. On arXiv, delve fell 77%, from 45.4 ppm to 10.4, while the list kept rising, from 5,612 to 5,858. The words that peaked first were the ones reported first.

When each era's words peaked on arXiv

Median peak month of the words in each registered group

  1. Oct 2023

    Words reported for 2023 to mid-2024

    Testament, garner, boast.
  2. Mar 2024

    The core five

    Delve, tapestry, meticulous, intricate, pivotal.
  3. Mar 2025

    Words reported for mid-2024 to mid-2025

    Fostering, showcasing, enhance, align with.
  4. May 2025

    Words reported from mid-2025

    Emphasizing, highlighting, multifaceted, prioritize.
The groups were fixed before counting; the three later ones follow Wikipedia's era lists. These are medians: the earliest and latest word in each group can be a year or more apart.

Source: host0 analysis of arXiv abstracts

2026 is when nearly every word on the list falls. The list on arXiv dropped from 6,533 ppm in March 2025, its monthly high, to 1,711 in October 2026. It fell in AI-type categories, other CS, physics and math alike.

The list of 30 on arXiv

Uses per million words in abstracts, monthly, Jan 2019 to Oct 2026

The list of 30 on arXiv
xy
Jan 20191,889
Feb 20191,981
Mar 20192,016
Apr 20192,008
May 20192,110
Jun 20192,218
Jul 20192,037
Aug 20192,040
Sep 20192,143
Oct 20192,118
Nov 20192,174
Dec 20192,168
Jan 20202,058
Feb 20202,132
Mar 20202,219
Apr 20202,204
May 20202,196
Jun 20202,336
Jul 20202,290
Aug 20202,288
Sep 20202,264
Oct 20202,380
Nov 20202,325
Dec 20202,347
Jan 20212,251
Feb 20212,329
Mar 20212,470
Apr 20212,467
May 20212,370
Jun 20212,446
Jul 20212,321
Aug 20212,391
Sep 20212,495
Oct 20212,482
Nov 20212,423
Dec 20212,411
Jan 20222,433
Feb 20222,452
Mar 20222,618
Apr 20222,555
May 20222,561
Jun 20222,595
Jul 20222,560
Aug 20222,550
Sep 20222,607
Oct 20222,755
Nov 20222,728
Dec 20222,584
Jan 20232,637
Feb 20232,773
Mar 20232,959
Apr 20233,157
May 20233,551
Jun 20233,653
Jul 20233,562
Aug 20234,130
Sep 20234,113
Oct 20234,422
Nov 20234,692
Dec 20234,737
Jan 20244,759
Feb 20245,086
Mar 20245,452
Apr 20245,321
May 20245,344
Jun 20245,623
Jul 20245,677
Aug 20245,856
Sep 20245,869
Oct 20245,978
Nov 20245,922
Dec 20246,261
Jan 20256,155
Feb 20256,495
Mar 20256,533 (6,533 · Mar 2025)
Apr 20256,200
May 20256,397
Jun 20256,204
Jul 20255,879
Aug 20256,019
Sep 20255,583
Oct 20255,269
Nov 20255,151
Dec 20254,743
Jan 20264,723
Feb 20264,398
Mar 20264,203
Apr 20263,920
May 20263,290
Jun 20263,090
Jul 20262,691
Aug 20262,308
Sep 20261,978
Oct 20261,711
The climb ran from early 2023 to March 2025, and the fall has taken about 19 months. By period: 2,335 in the baseline, 5,858 in 2025, 1,969 in Aug to Oct 2026. The same 30 words count in every month.

Source: host0 analysis of 1.75M arXiv abstracts

Put the two lines on one chart and the swap is easy to see. Delve turned in early 2024. The list carried on for another year.

Delve and the list of 30 on arXiv

Each as a multiple of its own pre-ChatGPT rate, centred 3-month rate, log scale, Jan 2022 to Oct 2026

Delve and the list of 30 on arXiv
Seriesxy
DelveJan 20221.66×
DelveFeb 20221.64×
DelveMar 20221.65×
DelveApr 20221.33×
DelveMay 20221.12×
DelveJun 20221.11×
DelveJul 20221.3×
DelveAug 20221.82×
DelveSep 20221.92×
DelveOct 20221.99×
DelveNov 20221.63×
DelveDec 20221.84×
DelveJan 20231.94×
DelveFeb 20232.37×
DelveMar 20232.83×
DelveApr 20233.72×
DelveMay 20235.28×
DelveJun 20236.18×
DelveJul 20238.77×
DelveAug 202311.44×
DelveSep 202314.77×
DelveOct 202317.17×
DelveNov 202319.23×
DelveDec 202320.8×
DelveJan 202420.84× (20.8× · Jan 2024)
DelveFeb 202420.58×
DelveMar 202420.05×
DelveApr 202419.11×
DelveMay 202416.83×
DelveJun 202415.9×
DelveJul 202414.58×
DelveAug 202413.46×
DelveSep 202411.1×
DelveOct 20249.79×
DelveNov 20249.29×
DelveDec 20248.29×
DelveJan 20257.31×
DelveFeb 20255.69×
DelveMar 20255.18×
DelveApr 20254.53×
DelveMay 20254.38×
DelveJun 20253.53×
DelveJul 20252.95×
DelveAug 20251.98×
DelveSep 20251.92×
DelveOct 20251.85×
DelveNov 20251.86×
DelveDec 20251.59×
DelveJan 20261.34×
DelveFeb 20261.24×
DelveMar 20260.8×
DelveApr 20260.66×
DelveMay 20260.6×
DelveJun 20260.68×
DelveJul 20260.61×
DelveAug 20260.38×
DelveSep 20260.45×
List of 30Jan 20221.04×
List of 30Feb 20221.08×
List of 30Mar 20221.09×
List of 30Apr 20221.1×
List of 30May 20221.1×
List of 30Jun 20221.1×
List of 30Jul 20221.1×
List of 30Aug 20221.1×
List of 30Sep 20221.13×
List of 30Oct 20221.16×
List of 30Nov 20221.15×
List of 30Dec 20221.14×
List of 30Jan 20231.14×
List of 30Feb 20231.2×
List of 30Mar 20231.27×
List of 30Apr 20231.39×
List of 30May 20231.49×
List of 30Jun 20231.54×
List of 30Jul 20231.62×
List of 30Aug 20231.69×
List of 30Sep 20231.81×
List of 30Oct 20231.89×
List of 30Nov 20231.97×
List of 30Dec 20232.03×
List of 30Jan 20242.08×
List of 30Feb 20242.19×
List of 30Mar 20242.27×
List of 30Apr 20242.3×
List of 30May 20242.32×
List of 30Jun 20242.38×
List of 30Jul 20242.45×
List of 30Aug 20242.48×
List of 30Sep 20242.53×
List of 30Oct 20242.54×
List of 30Nov 20242.59×
List of 30Dec 20242.62×
List of 30Jan 20252.7×
List of 30Feb 20252.74×
List of 30Mar 20252.75× (2.75× · Mar 2025)
List of 30Apr 20252.73×
List of 30May 20252.69×
List of 30Jun 20252.64×
List of 30Jul 20252.58×
List of 30Aug 20252.49×
List of 30Sep 20252.4×
List of 30Oct 20252.29×
List of 30Nov 20252.17×
List of 30Dec 20252.09×
List of 30Jan 20261.98×
List of 30Feb 20261.89×
List of 30Mar 20261.78×
List of 30Apr 20261.62×
List of 30May 20261.46×
List of 30Jun 20261.3×
List of 30Jul 20261.16×
List of 30Aug 20260.98×
List of 30Sep 20260.84×
Each gridline is ten times the one below it. Delve peaks at 20.8× around January 2024 and ends at 0.45×. The list peaks at 2.75× around March 2025 and ends at 0.84×.

Source: host0 analysis of arXiv abstracts

Word by word, on arXiv: "significant" went from 1,611 ppm to 462, now 0.58× its baseline. "Crucial" went from 427 to 102, or 0.59×. "Enhance" went from 1,232 to 273, or 0.93×. Geng and Trotta reported that "significant" kept rising in their data through 2024. That's right for their window: it kept rising until about February 2025, and this is what happened next.

Not everything has come all the way back. "Align with" is still at 3.3× on arXiv, "prioritize" at 3.1× and "underscore" at 4.6×, though all three are off their own peaks.

How many words peaked and fell, by the registered rule

Share of words whose latest rate is at most 2/3 of their peak, with 95% intervals apart

How many words peaked and fell, by the registered rule
LabelValue
arXiv, markers34 of 40
arXiv, controls8 of 30
Hacker News, markers15 of 40
Hacker News, controls1 of 30
Show HN, markers25 of 40
Show HN, controls2 of 30
GitHub, markers15 of 40 (23 inconclusive)
GitHub, controls1 of 30
The 40 markers are the 31 registered words, 8 phrases and the em dash; the 30 controls are ordinary words with no known LLM link. Every word is scored, none selected. Low-count words often land in neither bucket, which is the rule working. By the same rule, words still rising: arXiv 3 markers and 13 controls, Hacker News 4 and 10, Show HN 2 and 18, GitHub 2 and 16.

Source: host0 analysis

Why did it fall? We can't say. Model updates, people editing out known tells and fear of detectors all fit, and our data can't separate them. Geng and Trotta read the early drop as writers adapting: choosing among outputs or changing them. That's their reading, of their data through 2024. Nothing here measures how much anyone uses AI.

The controls moved too#

We counted 30 ordinary word families alongside the markers, in three frequency bands, to see how much a word moves when nothing in particular is happening to it. On Hacker News the answer is reassuring: the median control is at 1.01×, with a range of 0.82× to 1.34×. On arXiv it isn't. The median is 1.11×, but the range runs from 0.32× to 3.2×.

Ordinary words on arXiv moved as far as the markers did

Latest rate (Aug to Oct 2026) as a multiple of the pre-ChatGPT rate, selected controls with delve and the list

Ordinary words on arXiv moved as far as the markers did
LabelValue
never3.20×
remember3.07×
cheap2.81×
before2.65×
because1.47×
List of 300.84×
Delve0.45×
somewhat0.37×
obvious0.36×
puzzled0.32×
Controls shown are the extremes plus 'because'; all 30 are in our data. The list's 0.84× sits inside the control range, so on arXiv it doesn't clear a strict placebo. Hacker News controls stay within 0.82× to 1.34×.

Source: host0 analysis of arXiv abstracts

So arXiv's general vocabulary shifted in 2026, plain words included, and we have to say so plainly: the list's return to 0.84× doesn't clear a strict placebo on arXiv. A word like "before" moved further than the list did. These do clear it: delve's rise (20.8× at the peak, against a control maximum of 4.5×), delve's fall (latest over peak of 0.021, against 0.123 for the lowest control) and the rise of "genuinely". A March 2026 paper by Geng, Dong and Poibeau also reports shifts in ordinary function words on arXiv, such as "beyond" and "via" rising in titles and "the" and "of" falling in abstracts, which they consider likely to be LLM-driven. Abstracts have also grown, from about 162 words in our baseline to 188 in October 2026. We don't know why the plain words moved, so we list it under open questions.

Who says it on Hacker News#

It isn't a few accounts or an obvious bot farm. In every year, the top 10 authors hold 4 to 9% of "delve" uses and the top 1% hold 3 to 6%. A control word with a similar count, "clumsy", looks the same (the top 10 hold 6 to 11%). We looked at these only as aggregates, and our data contains no usernames.

Newer accounts do over-index, and the gap is shrinking. This cut is exploratory: we added it after seeing results.

Share of Hacker News comments from accounts under 90 days old

Comments containing "delve" against all comments, by year, 2019 to Oct 2026

Share of Hacker News comments from accounts under 90 days old
Seriesxy
Comments with "delve"20199.1%
Comments with "delve"20208.1%
Comments with "delve"20218.4%
Comments with "delve"20228.7%
Comments with "delve"202321.5% (21.5%)
Comments with "delve"202415.3%
Comments with "delve"202512.2%
Comments with "delve"202611.5%
All comments20198.1%
All comments20207.8%
All comments20217.5%
All comments20227%
All comments20236.9%
All comments20246%
All comments20256.8%
All comments20268.1%
Exploratory. The gap opened in 2023 (21.5% against 6.9%) and has closed to 11.5% against 8.1%. 'All comments' is a 1-in-20 sample. A new account isn't evidence about any one comment.

Source: host0 analysis of Hacker News comments

Older accounts moved too, in another exploratory cut using 1 author in 5. "Genuinely" among accounts that were already active before ChatGPT went from 41.4 ppm in 2021 to 76.9 in July to October 2026. Among newer accounts it's 92.1. So it isn't only new accounts.

Hacker News changed its rules in March 2026#

Anything that falls on Hacker News after March 2026 comes with a caveat. The site's guidelines now say "Don't post generated text or AI-edited text." Reading the moderators' own statements in the public dataset, they said on 8 March that Show HN would be restricted for new accounts for a while, and on 11 March that the rule on generated comments had moved from moderation practice into the written guidelines, and that accounts breaking it were being banned. By their account the rule itself wasn't new; writing it down and limiting Show HN were. We link the thread and name nobody.

Dead comments on Hacker News

Share of all comments marked dead, monthly, Jan 2024 to Oct 2026

Dead comments on Hacker News
xy
Jan 20242%
Feb 20242%
Mar 20242.3%
Apr 20244.4%
May 20242.1%
Jun 20242.3%
Jul 20242.7%
Aug 20242.6%
Sep 20242.6%
Oct 20242.1%
Nov 20242.6%
Dec 20242.6%
Jan 20252.4%
Feb 20252.8%
Mar 20253.2%
Apr 20252.8%
May 20252.7%
Jun 20252.7%
Jul 20252.5%
Aug 20252.2%
Sep 20252.8%
Oct 20252.3%
Nov 20252.4%
Dec 20252.2%
Jan 20262.8%
Feb 20263.4%
Mar 20265.6% (5.6% · Mar 2026)
Apr 20265.8%
May 20265.6%
Jun 20265.4%
Jul 20265.4%
Aug 20265.7%
Sep 20265.6%
Oct 20266.2%
2.57% across 2025 (85,490 of 3.32M), then 5.67% from March to October 2026 (151,339 of 2.67M). The one-month jump in April 2024 is one we didn't look into. Hacker News has no text for dead comments from about mid-2022, so we can only count them.

Source: host0 analysis of Hacker News comments

Em dashes on Hacker News went up and then back down, with a peak just before the rule took effect.

Em dashes in Hacker News comments

Uses per million words, monthly, Jan 2022 to Oct 2026

Em dashes in Hacker News comments
xy
Jan 2022215
Feb 2022215
Mar 2022208
Apr 2022231
May 2022217
Jun 2022229
Jul 2022231
Aug 2022207
Sep 2022211
Oct 2022218
Nov 2022220
Dec 2022228
Jan 2023219
Feb 2023218
Mar 2023234
Apr 2023236
May 2023247
Jun 2023226
Jul 2023237
Aug 2023220
Sep 2023235
Oct 2023233
Nov 2023237
Dec 2023230
Jan 2024248
Feb 2024244
Mar 2024240
Apr 2024250
May 2024269
Jun 2024252
Jul 2024258
Aug 2024263
Sep 2024283
Oct 2024292
Nov 2024313
Dec 2024340
Jan 2025309
Feb 2025367
Mar 2025395
Apr 2025457
May 2025450
Jun 2025433
Jul 2025461
Aug 2025413
Sep 2025366
Oct 2025383
Nov 2025361
Dec 2025368
Jan 2026359
Feb 2026589 (589 · Feb 2026)
Mar 2026537
Apr 2026379
May 2026285
Jun 2026273
Jul 2026292
Aug 2026281
Sep 2026251
Oct 2026211
The baseline is 220 ppm, so plenty of people typed them before ChatGPT. The centred 3-month peak is 503 (Feb to Apr 2026). October 2026 is 211, under the baseline. Aug to Oct 2026 is 247, or 1.12×.

Source: host0 analysis of Hacker News comments

Show HN fell harder. Em dashes there went from 3,491 ppm in the first quarter of 2026 (6,245 in 1.79M words) to 244 in October, as posts went from 2,408 in October 2025 to 6,215 in February 2026 and 2,641 in May, then back to 4,241 in October, and the share of posts with body text fell from about 70% to 38%.

So a fall on Hacker News after March 2026 may be removal rather than changed writing, and we can't separate the two. The timing lines up. We can't say the rule caused it, because the comments that were removed can't be read.

Where the tell went: the em dash and genuinely#

Everything in this section is exploratory. The em dash and "genuinely" are both in our registered list, but the README document share and the decision to single these two out came after we'd seen results. Treat them as leads.

Start with the em dash (—). In our sample of READMEs of GitHub repos created from January to October 2026, 39.0% contain one: 841 of 2,156, with a 95% interval of 37.0% to 41.1%. For repos created in 2024 it's 1.9% (38 of 1,984), and for 2022 it's 0.6% (11 of 1,734). "New" here means each README as it stood 30 days after the repo was created. This isn't a rate for GitHub as a whole, because it pools two sampling strata: all repos (34.6% in 2026, 1.8% in 2024) and repos with at least one star (44.5% in 2026, 2.0% in 2024).

Sampled READMEs that contain an em dash

By the year the repo was created; each README as it stood 30 days after creation; 2026 is Jan to Oct

Sampled READMEs that contain an em dash
LabelValue
20192.1% (87 of 4,099)
20201.1% (17 of 1,543)
20211.1% (19 of 1,742)
20220.6% (11 of 1,734)
20231.5% (28 of 1,890)
20241.9% (38 of 1,984)
202511.7% (245 of 2,103)
202639.0% (841 of 2,156)
Exploratory. English READMEs of 30 words or more, deduplicated, pooled across two strata. In 2026, all repos are at 34.6% and repos with at least one star at 44.5%. 2019 has more repos only because the first sampling pass started there.

Source: host0 analysis of a sample of 17,251 GitHub READMEs

A few files don't drive this: in 2026, the top 1% of READMEs hold 19.5% of the dashes (2024: 81.4%), and 57 of the 841 use one as an empty table cell. READMEs did get longer, from a median of 129 words in 2022 to 318 in 2026, and a longer document has more chances to contain a dash. But the rate per word rose too, from 174 ppm in the baseline to 6,931 in July to October 2026. Held at 2024's mix of lengths, the 2026 share is 30.0% rather than 39.0%, so length explains part of the level and little of the rise. The pooled figure also hides a climb through the year: 17.9% for January's repos, 50.8% for June's and 52.6% for August's.

Why READMEs, when the em dash on Hacker News sits at 1.12×? We have reasons to be careful, and one guess.

  • Who makes repos. We can't separate a change in who creates repos from a change in how the same people write. New people, and agents creating repos for them, would look the same here. Some simple cuts don't explain it away: at 2024's mix of length, language and stratum the 2026 share is 29.0%, with one repo per owner it's 38.9%, and only 20 of the 841 share a dash-bearing line with another owner's README, so it isn't one template. But the frame moved as well. We list about 146 surviving repos a minute for 2026 against about 86 for 2024 (older cohorts have had longer to lose repos), and more of what we read passes our filters: 25% of all repos against 21%, and 58% of starred repos against 41%. Counted against every repo we read, with or without a usable README, about 17.0% have a dash for 2026 and 0.7% for 2024, so 39.0% is a level for READMEs that pass our filters.
  • Survivorship. We can only read repos that still exist and are public. A random GH Archive sample, which includes repos deleted later, shows 32 to 42% of each cohort gone, with no steep trend by age. That frame's list rate agrees with ours (2,684 ppm in 2025 against 2,682), though its em dash counts are too small to use. The check stops at 2025 cohorts, because GH Archive's repo events stop in October 2025, so 2026 has no check, and younger repos have had less time to disappear.
  • A time-boxed sample. About 100 minutes of collection listed 140,747 repos, read 71,178 and kept 17,251 READMEs, or 5.4M words. Most drop-outs have no README or too few words. That's enough for lists and document shares, and not enough for rare single words, so we cite none on GitHub beyond the raw count for "genuinely" below. Repos created in October 2026 have under 30 days of history.
  • Dating. We read each README at 30 days because current text with a creation date would backdate rewrites. On the same repos, em dashes at 30 days against today: 2 → 14 for 2024 repos, 78 → 218 for 2025 repos, 13 → 45 for 2019 repos. So old READMEs get rewritten, and our rule keeps that out.
  • Our guess: nobody edits a README the way people edit an abstract or a comment. There are no coauthors, no moderators and no reviewers. We haven't measured this.

Louis Abraham's separate study of GitHub pull request descriptions points the same way: its README gives about 0.2 em dashes per 10,000 words in early 2024 and 123 in mid-2026, in a corpus of 461,121 descriptions. The study reads vocabulary only and says it can't tell who wrote a text. We leave arXiv out of this: abstracts are almost entirely ASCII, authors type ---, and that count swings for reasons unrelated to style. And the em dash proves nothing about any single author. In 2021, one Hacker News account alone wrote more than a thousand of them.

The tell may already be moving again, but the evidence is thin. On READMEs the em dash peaked at 10,486 ppm in the second quarter of 2026. The third quarter is 10,350, so nearly all of the fall to 6,931 for July to October comes from October (3,130), whose repos have under 30 days of history. By document share, September's repos (41.1%, only 73 READMEs) and October's (32.9%) sit under August's 52.6%.

"Genuinely" against its pre-ChatGPT rate

Rate as a multiple of the Jan 2019 to Oct 2022 rate, by year, then Aug to Oct 2026

"Genuinely" against its pre-ChatGPT rate
Seriesxy
arXiv abstracts20221.08×
arXiv abstracts20231.04×
arXiv abstracts20241.42×
arXiv abstracts20252.26×
arXiv abstractsSep 20266.15×
Hacker News comments20221.06×
Hacker News comments20231.08×
Hacker News comments20241.13×
Hacker News comments20251.49×
Hacker News commentsSep 20261.94×
Exploratory. arXiv 3.71 → 22.81 ppm (401 → 484 uses), above the 3.2× control maximum. Hacker News 39.62 → 76.81 ppm (30,019 → 3,916 uses), above the 1.34× control maximum. In a sample of 30 Hacker News uses, all 30 were the ordinary adverb.

Source: host0 analysis of arXiv and Hacker News

"Genuinely" is the one registered word that's up and clear of the controls on arXiv, Hacker News and READMEs and isn't a topic word. On READMEs the counts are tiny: from 3 uses to 29, so we don't quote a multiple. It's also starting to slip: on arXiv from 27.3 ppm (June to August 2026) to 19.0 in October, and on Hacker News from 84.0 (February to April 2026) to 73.8 in October. "Align with" is up and clear on arXiv and Hacker News as well, but not on READMEs.

Some words that Abraham's study surfaced, as Implicator reported in August 2026, also show up in arXiv abstracts, a separate corpus. Baseline to latest, in ppm: byte-identical 0.01 → 2.40 (1 → 51 uses), load-bearing 0.48 → 3.53, quietly 0.11 → 1.55, seam 1.44 → 6.64, survived 5.09 → 36.2. We added these after registering, so they're exploratory, and remember arXiv's plain words moved as much as 3.2×.

Words move for reasons that have nothing to do with AI#

A word count can't read. Here's what it picked up along the way.

A startup called Delve. Hacker News's raw "delve" rate jumps to 8.2 and 8.8 ppm in March and April 2026, and April is the highest month in our data. The cause is a compliance startup with that name, backed by Y Combinator at the time. TechCrunch reported allegations against it in March, which Delve disputes, and in April that the two had parted ways. Title-case uses mid-sentence go from usually under 5 a month to 51 and 77.

"Delve" on Hacker News, with and without the noise

Uses per million words, monthly, Jul 2024 to Oct 2026

"Delve" on Hacker News, with and without the noise
Seriesxy
All usesJul 20246.18
All usesAug 20245.53
All usesSep 20247.88
All usesOct 20245.51
All usesNov 20244.29
All usesDec 20245.54
All usesJan 20253.08
All usesFeb 20253.8
All usesMar 20255.63
All usesApr 20253.47
All usesMay 20254.35
All usesJun 20256.03
All usesJul 20253.88
All usesAug 20255.2
All usesSep 20252.87
All usesOct 20252.69
All usesNov 20253.97
All usesDec 20253.03
All usesJan 20263.4
All usesFeb 20263.2
All usesMar 20268.17
All usesApr 20268.77 (8.77 · Apr 2026)
All usesMay 20262.98
All usesJun 20262.63
All usesJul 20263.98
All usesAug 20263.47
All usesSep 20262.72
All usesOct 20261.65
Without names and talk about the wordJul 20244.24
Without names and talk about the wordAug 20242.8
Without names and talk about the wordSep 20243.98
Without names and talk about the wordOct 20243.4
Without names and talk about the wordNov 20242.42
Without names and talk about the wordDec 20243
Without names and talk about the wordJan 20251.82
Without names and talk about the wordFeb 20252.49
Without names and talk about the wordMar 20253.49
Without names and talk about the wordApr 20252.2
Without names and talk about the wordMay 20252.33
Without names and talk about the wordJun 20252.29
Without names and talk about the wordJul 20251.59
Without names and talk about the wordAug 20252.54
Without names and talk about the wordSep 20251.74
Without names and talk about the wordOct 20251.85
Without names and talk about the wordNov 20252.05
Without names and talk about the wordDec 20251.6
Without names and talk about the wordJan 20261.81
Without names and talk about the wordFeb 20262.11
Without names and talk about the wordMar 20261.99
Without names and talk about the wordApr 20261.89
Without names and talk about the wordMay 20261.19
Without names and talk about the wordJun 20260.88
Without names and talk about the wordJul 20261.54
Without names and talk about the wordAug 20261.61
Without names and talk about the wordSep 20261.45
Without names and talk about the wordOct 20261.18
The second line is a keyword filter over every comment, a rough proxy we checked by hand. It removes the startup's spike, as it should. Other names in the data include a Go debugger, a World of Warcraft mode and a line of Tolkien.

Source: host0 analysis of Hacker News comments

The Bayeux Tapestry. "Tapestry" is the one core word rising on Hacker News, from 0.39 to 1.86 ppm (95 uses). In a sample of 25 from August to October 2026, 13 are about the Bayeux Tapestry and 5 about the Star Trek episode. One is about the word. That's news, not a marker.

Load-bearing. An exploratory word. It went from 0.53 ppm on Hacker News to 24 in August 2026, but about 7 in 10 of a sample of 30 are people joking about it or discussing it as a model's habit. A word's fame shows up in the count of the word.

People talking about delve. We hand-checked random Hacker News comments containing the word. Before ChatGPT, 93.8% were ordinary use. In 2026 it's 52.5% (63 of 120, 95% interval 43.6% to 61.2%), with 18.3% about the word itself and 29.2% names and quotes. The 2026 sample includes the startup's spike: title-case mid-sentence uses in March and April alone are 128 of the year's 700, so part of that 29.2% is one news story.

How often "delve" on Hacker News is just the word

Share of hand-checked comments using it in the ordinary sense

How often "delve" on Hacker News is just the word
LabelValue
Before ChatGPT93.8% (75 of 80)
202396.3% (77 of 80)
202474.2% (89 of 120)
202576.2% (61 of 80)
202652.5% (63 of 120)
Adjusted for this, ordinary delve went from 2.87 ppm before ChatGPT to 4.52 in 2024 and 2.15 in 2026 (1.79 to 2.50). A keyword proxy over every comment agrees on the direction: 2.58, 3.67, 1.57. Sample sizes are small.

Source: host0 hand check of random Hacker News comments

What this means if you build with AI#

  1. Don't use a word list as a detector. A list like ours would score a 2026 arXiv abstract about the same as a 2019 one (0.84× against 0.89×). Human text has these words and AI text doesn't need them, so it fails in both directions. Liang and colleagues found that GPT detectors consistently misclassify non-native English writing as AI-generated, which is one more reason not to build on this.
  2. Assume any tell has a short shelf life. Median peaks on arXiv ran from October 2023 to May 2025 across eras. "Genuinely" has already started to slip. If you keep a list for review or moderation, date it and expect to replace it.
  3. Don't judge a person by a word. One account wrote more than a thousand em dashes in 2021, and in a hand check of 30 Hacker News comments every "genuinely" was ordinary use. Ask about the work instead.
  4. Read the README your agent wrote. Before you publish it, run each command, check each claim and cut what you wouldn't say yourself. The em dash isn't the problem. A README nobody checked is. In our sample, 39.0% of READMEs of repos created from January to October 2026 have a dash at 30 days old.
  5. Ask for facts, not a banned-word list. Telling an agent to avoid "delve" changes the word and nothing else: one README in our 2026 sample is itself a prompt telling a model to avoid "delve into". Ask for what the reader needs: what it does, how to run it, what doesn't work yet.
  6. If you measure vocabulary, fix the list first and keep controls. Without controls we'd have reported arXiv's 0.84× with more confidence than it deserves. Without field splits we wouldn't have seen that harness was a topic. Both changed what we can claim.

Agents write the app and its README in the same afternoon. We built host0 so that what they make gets a link you can share, private until you say otherwise, once you've read it.

Related: what happened to the repos behind the hype, in the half-life of hyped AI repos, and who posts their AI chats in public, in who links to AI chats?

Methodology#

All data is public, pulled on 5 November 2026. Every number stops at 31 October 2026.

  • Pre-registration. We wrote down the corpora, word list, controls, measures and decision rules before computing any marker or control count. That order rests on file times on our own machine, not on a third-party timestamp, so treat it as our account rather than proof. The executable list is scripts/patterns.py (sha256 starts 4ce95339, unchanged at the end). Our registration file lists the later additions to the analysis, and we mark those as exploratory in the post. The next bullet also covers what that file leaves out.
  • What departs from it. The list-level summaries and removing harness; the case split of "delve" and the words from Abraham's study; the README document shares, including the em dash; account-cohort and per-author tables on Hacker News; and picking out "genuinely". The moderators' statements and the monthly dead-comment and Show HN post counts also came after we saw results. We changed the GitHub frame's window lengths once, from 2 minutes to between 2 and 12 by year, and read part of the sample through a faster path until it was rate-limited. The registered dead-comment sensitivity failed because that text doesn't exist, so we only give counts. The registered relative index (the rate against the median control) isn't charted: for the list it's 0.76×, 1.00× and 0.73×. The registered sensitivity variants for pivotal, realm and underscore are in the data.
  • arXiv. 1,751,361 abstracts and 296.8M words, from January 2019 to October 2026, from a CC0 snapshot on Hugging Face plus an OAI-PMH top-up of 10,480 October 2026 papers. Month is the arXiv id. The text is the latest abstract version; single-version papers give the same result (0.86×).
  • Hacker News. 26.7M comments and 1.57B words from the public ClickHouse copy (hackernews_history), excluding deleted and dead comments, with HTML, code blocks and links removed. Show HN is stories whose title starts with "Show HN", title plus text, by quarter. A parity check over 5,729 comments gave the same counts for all 73 patterns in both our pipelines.
  • GitHub READMEs. Complete listings of surviving public non-forks created in random short windows, from GitHub search, in two strata: all repos, and repos with at least one star on random days. Each README is read at its last commit within 30 days of creation, stripped of code, HTML, URLs and badges, and kept at 30 words or more if it looks like English, with exact duplicates dropped within a month. The pooled numbers are a sample, not a GitHub-wide rate.
  • Measures. Word = [A-Za-z]+(?:['’][A-Za-z]+)?; rate = matches per million words. Baseline = January 2019 to October 2022. Peak = highest centred 3-month rate from November 2022 (highest quarter for Show HN and GitHub). Latest = August to October 2026 (July to October for Show HN and GitHub). Brackets are 95% cluster bootstrap intervals by author, paper or repo. "Peaked and fell" = latest at most 2/3 of peak, with intervals apart.
  • Controls. 30 ordinary word families in three frequency bands, all kept. A marker's move counts as a signal only if it's outside the range of the controls' moves.
  • Limits. These are word counts, not readings. Nothing here detects AI text, and we have no data on author language. HN's moderation can't be separated from changes in writing after March 2026, and on GitHub we can't separate who creates repos from how they write. Peaks are biased up a little. The Show HN index against baseline isn't usable, and neither are rare words on GitHub. Hand checks are small samples.

The data pack, with every source, the sensitivity tables and a do-not-cite list, is in our repository at docs/research/2026-11-05-delve-index.md. The series behind every chart, the registration and the scripts are in docs/research/2026-11-05-delve-index-data/. It holds no usernames, repo names or comment text.

Open questions#

  • Why did arXiv's plain words move in 2026? "Before", "never" and "because" are up across fields, including math, and abstracts got longer. That could be model style, editing tools or something else.
  • Are the em dash and "genuinely" next to fade? Both have turned down in the last few months on most platforms.
  • Is Hacker News's drop after March writing or removal? Dead-comment text isn't available to check.
  • How did non-native writers' word choice shift? We have no author-language data on any corpus.
  • What's behind arXiv's volume? October 2026 has 42,152 new papers against 27,692 in October 2025, up 52%, and math went from 3,337 in January 2024 to 8,817 in October 2026. We haven't looked into why.
ResearchAI writing