Skip to the content
Dead Writing Theory

The Court of Public Prose

The indictment · In re Dead Writing Theory

The theory before the court Since ChatGPT, the writing on the web has been collapsing into one machine voice.

Same words, same rhythm, same punctuation, everywhere.

  1. Count ISame wordsThe vocabulary has converged on one set of marketing words.3 P-1, P-2, P-36 D-1, D-4, D-6, D-7, D-8, D-9
  2. Count IISame rhythmThe sentences have converged on one set of shapes.0 No exhibits yet2 D-3, D-5
  3. Count IIISame punctuationThe punctuation has converged on one mark: the em dash.1 P-41 D-2

Exhibits entered against each count so far, prosecution and defence. 21 exhibits across 7 cases; 8 bear on the theory as a whole (how much writing is machine-made, how alike it is) rather than on one count: prosecution P-5, P-6, P-7, P-8, P-9, P-10, defence D-10, D-11.

§ IWhere the name comes from

Dead Writing Theory, after the Dead Internet Theory

Here is the theory, stated plainly. Since ChatGPT arrived, the writing on the web has been collapsing into one machine voice: same words, same rhythm, same punctuation, everywhere. Prose, copy, advertising, marketing. All of it drifting towards a single, slightly shiny, faintly confident hum.

You might recognise the shape of it. It's a nod to the Dead Internet Theory, the idea that much of the web is now made by machines, for machines. We liked the name so much we borrowed the grammar. But this site takes only the narrower, measurable corner: the writing itself. Forget the bots and the traffic, and whoever's lurking in the comments. Just the words on the page, and whether they've started to sound alike.

§ IIThe case for, the case against

Prosecution v. Defence

Both opening statements in full, then every exhibit so far, grouped by the table it was entered from. Any tally counts the evidence items below; it is not a measure of truth.

For the prosecution

The case for the prosecution

The prosecution would like to begin with a list of words you already know in your bones. "Seamless". "Unlock". "Streamline". After ChatGPT, these AI-era marketing words nearly doubled across company homepages, from 1.05 to 1.93 uses per 1,000 words. And it wasn't a few loud outliers: the words rose on 83 sites and fell on only 27. That's a pattern, members of the jury, not a fluke.

Exhibit P-2 is one word in particular. "Unlock" turned up 4.4 times as often in 2023 to 2026 as in 2019 to 2022, going from 7.1 to 31.3 uses per 100,000 words. Everybody, apparently, had a door to open.

Exhibit P-3 is the one that stings. In 2023, the AI-era words overtook the dotcom words ("solutions", "world-class", "click here") for the first time in 27 years. A whole generation of corporate filler was quietly replaced by its successor, and nobody held a funeral.

Finally, the punctuation. In 2023 to 2026, 69% of software homepages use an em dash, against 17% of banking and finance homepages. The industry closest to the machines writes most like them. The prosecution rests, and it would like you to notice the dash count in this very sentence. (There isn't one. We checked.)

For the defence

The case for the defence

The defence will be brief on adjectives and heavy on dates. Start with the timeline. The rise in those marketing words began before ChatGPT: 0.45 uses per 1,000 words in 2012, 1.29 in 2019. You can't blame the newcomer for a drift that was already well under way. It just turned up late and took the credit.

Next, the em dash. It was on 3% of homepages in 2000, 27% in 2022 and 38% in 2026. That's a long, steady climb, and most of it happened before ChatGPT existed.

Then the famous sentence shapes: "not X but Y", the reflexive list of three, "furthermore". The things people circle in red. They did not increase. They ran at 0.16 per 1,000 words before and 0.11 after. If anything, they dipped.

Now widen the lens. Before ChatGPT existed, press releases used AI-isms 3.7 times as often as news did: 1.61 against 0.44 per 1,000 words. The "AI" sentence shapes were most frequent in pre-1928 speeches (2.07 per 1,000 words) and sermons (1.94), against 0.45 in UK news. And 40 of 43 words now called AI-isms were more common in books in 2019 than in 1980.

Last, the "delve" story. Said to have come from Nigerian English, it didn't hold up: 3 of 600 Nigerian news pages used "delve", 1 of 600 Kenyan, 0 of 1,000 UK and US. The defence rests. Quietly. With a small, smug cough.

Evidence for the theory

For the prosecution

10 exhibits
  1. Exhibit P-1Case No. DWT-001

    1.05 → 1.93

    uses per 1,000 words

    AI-era marketing words nearly doubled after ChatGPT

    Each of 147 company homepages compared with itself, 2019 to 2022 against 2023 to 2026. Up on 83 sites, down on 27.

    Entered for the prosecution · ScriptGrain SGR-004

  2. Exhibit P-2Case No. DWT-001

    4.4×

    7.1 → 31.3 uses per 100,000 words

    "Unlock" appeared 4.4 times as often

    Company homepages, 2023 to 2026 against 2019 to 2022.

    Entered for the prosecution · ScriptGrain SGR-004

  3. Exhibit P-3Case No. DWT-001

    2023

    first year

    In 2023 the AI-era words overtook the dotcom words for the first time in 27 years

    "Seamless", "unlock" and "streamline" passed "solutions", "world-class" and "click here".

    Entered for the prosecution · ScriptGrain SGR-004

  4. Exhibit P-4Case No. DWT-001

    69%

    of software homepages use an em dash (2023 to 2026)

    Software homepages are the most converged

    Against 17% of banking and finance homepages. Software also carries the most AI-era vocabulary: 3.88 uses per 1,000 words, against 0.39 for insurers.

    Entered for the prosecution · ScriptGrain SGR-004

  5. Exhibit P-5Case No. DWT-003

    24%

    of corporate press releases, late 2024

    About a quarter of corporate press releases involved LLM-assisted writing

    537,413 corporate press releases, January 2022 to September 2024. A population-level estimate from word distributions, not a per-document detector.

    Entered for the prosecution · Liang et al., Patterns 6(12), 2025 · third-party study

  6. Exhibit P-6Case No. DWT-003

    18%

    of financial consumer complaints, late 2024

    Financial consumer complaints took on the machine's words too

    Roughly 18% of 687,241 consumer complaints. Also nearly 10% of job postings in small firms and 14% of UN press releases.

    Entered for the prosecution · Liang et al., Patterns 6(12), 2025 · third-party study

  7. Exhibit P-7Case No. DWT-005

    2.1% → 29.4%

    of sites, second half of 2022 → first half of 2025

    LLM-dominant websites rose steadily

    Roughly 100k sites archived by Common Crawl, classified site by site. A preprint; on pre-ChatGPT sites the classifier flagged only 0.29%.

    Entered for the prosecution · He et al., arXiv preprint 2605.00087, 2026 · third-party study

  8. Exhibit P-8Case No. DWT-005

    46.6%

    of 10,000 how-to searches had one in the top 10

    Machine-written sites reach the top of how-to searches

    65.7% for the top 20 results. 78.8% of LLM-dominant sites have a clear financial incentive, against 55.8% of other sites.

    Entered for the prosecution · He et al., arXiv preprint 2605.00087, 2026 · third-party study

  9. Exhibit P-9Case No. DWT-006

    50.9%

    of new English articles primarily AI-generated, Q4 2025

    Machine-written articles drew level with human ones

    49.6% against 50.4% human in Q1 2025. Vendor study: 55.4k Common Crawl articles, three commercial detectors averaged.

    Entered for the prosecution · Paredes et al., Graphite research (Five Percent), 2026 · third-party study

  10. Exhibit P-10Case No. DWT-007

    10.7%

    of the total similarity range, with one AI idea

    Stories written with AI ideas grew more alike

    8.9% with up to five AI ideas. 293 writers in a preregistered experiment; similarity measured with text embeddings.

    Entered for the prosecution · Doshi and Hauser, Science Advances, 2024 · third-party study

Evidence against the theory

For the defence

11 exhibits
  1. Exhibit D-1Case No. DWT-001

    0.45 → 1.29

    uses per 1,000 words, 2012 to 2019

    The rise started long before ChatGPT

    AI-era vocabulary was already climbing three years before ChatGPT launched.

    Entered for the defence · ScriptGrain SGR-004

  2. Exhibit D-2Case No. DWT-001

    3% → 27%

    of homepages, 2000 to 2022

    The em dash was spreading for two decades

    38% by 2026. The rise after ChatGPT continues a line that started in 2000.

    Entered for the defence · ScriptGrain SGR-004

  3. Exhibit D-3Case No. DWT-001

    0.16 → 0.11

    AI-style sentence shapes per 1,000 words

    The sentences did not change

    "Not X but Y", reflexive lists of three, "furthermore": no increase on company homepages after ChatGPT.

    Entered for the defence · ScriptGrain SGR-004

  4. Exhibit D-4Case No. DWT-002

    3.7×

    as many AI-isms as news, pre-ChatGPT

    Press releases wrote like this before any model did

    1.61 against 0.44 uses per 1,000 words in pages from 2021 and earlier.

    Entered for the defence · ScriptGrain SGR-007

  5. Exhibit D-5Case No. DWT-002

    2.07

    per 1,000 words in pre-1928 speeches

    The "AI sentence shapes" were most at home in old speeches and sermons

    Sermons and religion 1.94; UK news 0.45.

    Entered for the defence · ScriptGrain SGR-007

  6. Exhibit D-6Case No. DWT-002

    40 / 43

    more common in 2019 than 1980

    40 of 43 "AI-isms" were already rising in books

    Google Books. "Transformative" rose 31 times between 1980 and 2019.

    Entered for the defence · ScriptGrain SGR-007

  7. Exhibit D-7Case No. DWT-002

    3 / 600

    Nigerian news pages used "delve" (2021)

    The "delve came from Nigerian English" story did not hold up

    1 of 600 Kenyan, 0 of 1,000 UK and US. Rare everywhere; the difference was within chance.

    Entered for the defence · ScriptGrain SGR-007

  8. Exhibit D-8Case No. DWT-004

    214.45 → 254.65

    MTLD lexical diversity, 2018 → 2024

    Lexical diversity in news did not fall

    Two samples of roughly 30,000 news articles each. Machine-style words did rise, from 0.230% to 0.347% of words, but the homogenisation did not show.

    Entered for the defence · Fitterer et al., Proceedings of ACL, Student Research Workshop, 2025 · third-party study

  9. Exhibit D-9Case No. DWT-004

    0.00110

    change in MATTR lexical diversity, 2018 → 2024

    Two other diversity measures barely moved

    Against within-year variation of 0.00109 (MATTR 0.88011 → 0.88121). Maas moved from 0.01469 to 0.01482, a difference of 0.00013 against 0.00016.

    Entered for the defence · Fitterer et al., Proceedings of ACL, Student Research Workshop, 2025 · third-party study

  10. Exhibit D-10Case No. DWT-006

    49.9%

    of new articles primarily AI-generated, Q1 2026

    The machine share stopped rising

    "Since Q1 2025 the percentage of primarily AI-generated articles has plateaued at roughly 50%." Vendor study, detector-based.

    Entered for the defence · Paredes et al., Graphite research (Five Percent), 2026 · third-party study

  11. Exhibit D-11Case No. DWT-007

    8.1%

    rise in novelty with up to five AI ideas

    Each story got better with AI ideas

    Usefulness rose 9.0%. Less creative writers gained most: 10 to 11% on creativity, 22 to 26% on how enjoyable and well written the story is.

    Entered for the defence · Doshi and Hauser, Science Advances, 2024 · third-party study

§ IIIHow evidence is admitted

Counted, not guessed.

Every court needs rules, or it's just a pub argument with better chairs. Ours are short. In our own studies, the questions and the word lists are fixed before anything is measured, so nobody gets to go fishing for a result they like. Everything we measure is measured with deterministic code, never a language model. Same input, same answer, every time, with no vibes involved. Other people's studies are entered too, credited by name, with their methods and caveats read into the record.

Every case links to its full method, and to the original study. If you think we got it wrong, you can check, and we'd honestly rather you did. And both sides are always shown. The prosecution and the defence get their full say, even when one of them is clearly having a better afternoon.

  1. Rule I. The question is fixed first. The question is written down before anything is measured. Each case states it at the top, like a charge.
  2. Rule II. Counted by code, never by a model. Our own studies are counted with deterministic code that gives the same answer every run. No language model is asked for an opinion. A third-party study is entered with its authors' method stated and its caveats beside it.
  3. Rule III. Every exhibit goes in the record. Every finding is entered, for the prosecution or for the defence, and the full method and data are published on ScriptGrain, or, for a third-party study, linked at the original.

§ IVThe docket so far

7 verdicts. Three split verdicts, two for the prosecution and two for the defence.

The exhibits each case entered, side by side. A ledger, never one score: a count of exhibits is not a measure of truth.

Cases by verdict, with the exhibits each one entered
Case No.MatterProsecutionDefenceVerdict
DWT-00127 years of company homepages43Splitverdict
DWT-002Where the AI-isms came from04Verdict for the defence
DWT-003A quarter of the press release · Third-party study20Verdict for the prosecution
DWT-004The news that didn't flatten · Third-party study02Verdict for the defence
DWT-005The websites nobody wrote · Third-party study20Verdict for the prosecution
DWT-006Half the articles, and a plateau · Third-party study (vendor)11Splitverdict
DWT-007Better stories, more alike · Third-party study11Splitverdict
DWT-008Being measured––Pending, under investigation

§ VInstructions to the jury

Before you deliberate.

Is this the Dead Internet Theory?

Dead Writing Theory is a nod to the Dead Internet Theory: the idea that much of the web is now made by machines, for machines. We take the narrower, measurable corner of it, the writing: prose, copy, advertising and marketing. Not affiliated with the Dead Internet Theory or any site, forum or person associated with it.

Is this a real court?

No. The Court of Public Prose is a figure of speech for how we weigh the evidence: two sides, every exhibit in the record, a verdict per case. No court, agency or government is involved.

Do you use AI to measure the writing?

Not in our own studies: they are measured with deterministic code, never a language model, and the questions are fixed before anything is measured. Third-party studies are entered with their authors' own methods, some of which use AI-text detectors; each case says so in its caveats.

So is the web’s writing dying?

We don’t argue for the theory or against it. So far: 7 cases, three split verdicts, two for the prosecution and two for the defence. The jury is still out.

Who runs the studies?

ScriptGrain runs our own studies, and each of those cases links to its full method and data on scriptgrain.com. Third-party studies are credited to their authors and linked to the original. Ours: made by the same studio that runs this site.

Can I submit evidence?

Yes: hello@deadwritingtheory.com. The next case is being measured.