Corporate Reports Taught Chatbots How to Talk

Opening

Reader, these days a single long dash is enough to draw the comment “this was written by AI.” The mark English speakers call the em dash has quietly come to be treated as a machine’s fingerprint. But when The Economist lined up 1.2 million words of human and AI writing side by side and counted, it turned out ChatGPT now uses fewer dashes than human writers do. The hunters, it seems, are still carrying an outdated sketch.

So what’s the real fingerprint? And where did it come from? Let me give you the conclusion up front: the birthplace of today’s AI prose style isn’t AI at all — it’s human writing. Specifically, the kind of writing produced to be formally reviewed, like corporate reports and academic papers.

---

Nobody Noticed the Fingerprint Move

Let’s start with the experiment design. The Economist gave four models — ChatGPT, Claude, Gemini, and Grok — only a summary of one of its own articles and asked each to rewrite the piece on the same topic, without web search. It then compared the resulting human and machine writing across 55,940 sentences and 1.2 million words — roughly fifteen books’ worth of text. To rule out the possibility that its own house style was simply unusual, it set up control groups: the average of New York Times, Washington Post, and CNN articles from 2018-2022, and excerpts from bestselling novels published between 1950 and 2022.

The first finding is a changing of the guard among buzzwords. Delve and tapestry, once emblematic of AI writing, have all but vanished from the latest models. In their place, polysyllabic words like significant, increasingly, and consequences have moved in. In other words, catching AI by flagging a single telltale word has a shelf life measured in months.

The second finding is the dash reversal. Among the latest model versions, only Claude used more dashes than humans did. ChatGPT used the fewest of all the writers in the study — including the human ones.

So what is the real fingerprint, as of right now? It boils down to four things.

The vocabulary runs heavy. All four models use more eight-plus-letter words than humans do — rare words like interdependence, and scientific-sounding vocabulary like parameter or methodology. Nominalization¹, turning verbs into nouns, shows up clearly too: writing expansion where expand would do.

Punctuation disappears. Commas and semicolons appear less often than in human writing, and parentheses are nearly absent. Without expert quotes to cite, quotation marks are rare too.

Sentences run long, and the rhythm flattens out. The most overused word is the conjunction and. Sentence length varies less than in human writing, so there’s rarely a short sentence to break the flow.

Rhetorical formulas do the heavy lifting. Claude leans noticeably harder than humans on “not X, but Y” constructions, and ChatGPT overuses the rule of three.

One thread runs through all four: the absence of anyone else’s voice. No quotes, so no quotation marks. No interjected asides, so no parentheses. No reason to change pace, so sentence length stays even. It’s the statistical portrait of writing that talks to no one but itself.

And on top of all this sits one decisive chart. If you set the rate of typical AI-language use in The Economist’s own articles at 1, ChatGPT in 2024 measured around 6 — the latest models have come down to roughly 2 to 3. Each update moves closer to human writing. Which means the list I’ve just given you could be outdated in a few months.

I covered the problem of detectors chasing a moving target and catching innocent people in a previous issue. Today let’s follow the question in the opposite direction: where did this style come from in the first place?

---

The Spec Sheet Was Already Written in 1946

The Economist put a name to this list of vocabulary quirks: George Orwell’s “pretentious diction.”

In his 1946 essay “Politics and the English Language,” Orwell dissected the bad writing of his day. The symptoms he identified: dressing up simple statements in complicated words and jargon, stretching what a single verb could do into a noun phrase, vocabulary that puts on the costume of scientific impartiality, and the belief that words borrowed from Latin or Greek carry more prestige than plain Saxon ones.

Lay this list next to the AI fingerprint above. Long Latinate words match Orwell’s “dressing up.” Nominalization matches the noun-phrase habit he called “false limbs.” Scientific vocabulary matches the “costume of impartiality.” And indeed, in this analysis, all four models used more Latin-suffixed words than humans did. A spec sheet for bad writing drawn up 80 years ago is now doubling as the spec sheet for 2026’s AI prose.

There’s a part of Orwell’s diagnosis worth sitting with. He didn’t think this style was purely a product of laziness. When you have something uncomfortable to say, foggy words and the passive voice make a good shield. Nominalization is the classic move: the moment “I decided this” becomes “it was decided that this measure would be pursued,” the agent disappears from the sentence. That’s why formal style survives so stubbornly in government offices and corporations — not just because it sounds impressive, but because it’s useful for blurring accountability. The machine, it turns out, inherited an 80-year-old bureaucratic survival strategy, style and all.

Orwell didn’t stop at diagnosis — he left a prescription too. Of his six rules, three are worth quoting directly: “Never use a long word where a short one will do. If it is possible to cut a word out, always cut it out. Never use the passive where you can use the active.” The AI fingerprint above is, point for point, exactly what you get by breaking these three rules.

One fact worth flagging here: Orwell’s target was never machines. It was the writing of bureaucrats, politicians, and academics. The spec sheet came first; the machine that writes to it arrived 80 years later. So who exactly taught the machine this style?

---

We Went Looking for the Teacher and Found a Mirror

There are two suspects.

The first is the textbook — the training data. The text the models grew up reading is whatever has accumulated on the internet, and a large share of that isn’t everyday conversation but edited, formal written prose: papers, articles, reports, encyclopedias. Dashes, Latinate vocabulary, whatever — much of what gets called an “AI habit” is really just the statistical habit of formal written language.

But the textbook alone doesn’t explain everything. Florida State University’s Tom Juzek and Zina Ward pulled together 21 words — delve, intricate, underscore, and the like — that had surged in scientific-paper abstracts, then traced why exactly these words got overused. They found no clear cause in model architecture, in the algorithms, or in the training data. That leaves the second suspect, and the leading one: human feedback.

In a process called RLHF², human raters compare several answers a model has produced and pick the better one, and the model learns that preference. The question is what humans actually pick. In a judgment made in a few fleeting seconds, the writing that wins tends to be writing that sounds impressive rather than writing that’s accurate. Juzek’s explanation, as quoted by The Economist, points the same way: models pick up whatever people find impressive, and drop whatever people dislike.

The dash is exactly what shows how fast the “dropping” can happen. Once the dash became a punchline about AI, ChatGPT’s dash use plunged in the very next update. There’s a funny detail an Economist reporter noticed: earlier models, asked about overusing dashes, would banter back peppered with dashes of their own; the current model answers primly that it’s a mark to use sparingly. Style is tracking public opinion on a timescale of months.

One misconception worth clearing up: getting closer to human writing doesn’t mean becoming good writing. The Economist attaches the same caveat — the numerical gap has narrowed, but AI prose still lacks clarity and grace, and still leans on formulaic patterns. What statistics catch is a fingerprint, not the quality of a sentence. So as the fingerprint fades, the standard for judging writing swings back to plain craftsmanship. There’s one more barrier Juzek’s team flagged: since how these models are actually built isn’t disclosed, digging all the way down to root causes is itself difficult.

Now the picture comes together. Formal written language was fed in as the textbook. Impressive-sounding answers were rewarded. And whatever drew ridicule got deleted. At every one of these three stages, the standard was set by humans. AI style isn’t an invention; it’s a distillate — and the raw material is the writing we’ve been producing and approving all along. The chart showing AI writing edging closer to human writing with every update is, yes, a warning that detection will keep getting harder. But the thought that lingers longer is this: the distance between the two was probably never as great as we assumed.

---

Oswald’s Lens

I’ve spent more than 20 years writing, reading, and getting sign-off on strategy consulting documents. When I think back on the house language of that world, it goes like this: not “fix,” but “establish an improvement plan”; not “use,” but “enhance utilization.” In Korean documents, Sino-Korean nominalizations play the role that Latinate vocabulary plays in English. There’s an unspoken rule running through every approval chain: a short, native verb somehow looks low-status.

And from what I’ve observed, AI-generated Korean tilts in exactly that direction too: “through ~,” “from the ~ aspect,” “various,” “overall.” A good chunk of what people call “translation-ese” is, in fact, “report-ese.”

That’s why reading Orwell’s list stung a little. Before it was ever a spec sheet for AI style, it was a spec sheet for proposals I wrote years ago. The incentive that lets impressive-sounding writing clear approval existed first; AI simply distilled that incentive at scale. Which means the answer isn’t hunting for some trick list of “how to scrub the AI smell off your writing.” The direction runs the other way. Orwell’s principles were never meant as a way to dodge detection — they were guidance for good writing, and that’s still what’s needed now. Short words. Live verbs. If you can cut it, cut it.

---

Closing

To sum up:

1. The dash is an outdated sketch. Today’s fingerprint is heavy vocabulary, nominalization, vanishing punctuation, and rhetorical formulas.
2. This fingerprint overlaps with the spec sheet for bad writing Orwell drew up in 1946. It’s not a newly invented style.
3. The cause is a textbook made of formal written language, plus human feedback that rewarded writing that sounded impressive.

Try one experiment today. Take a document you’re working on and turn one nominalized phrase back into a verb — turn “an improvement plan needs to be established” into “I’ll fix it this way.” You’ll feel the sentence get shorter and the responsibility get clearer, right away.

If there’s a phrase at your company that survives purely because it “sounds impressive,” report it in the comments. Once enough examples come in, I’ll put together a “Dictionary of Korean Report-ese” in a future issue.

---

💬 Report the phrase that gets your documents approved, in the comments 📨 If you know a colleague who writes reports, forward this to them

---

References & Further Reading

Primary sources

• The Economist, “How to spot AI writing,” 2026. Link ··· This is the analysis that forms the backbone of today’s piece. Look at the original chart comparing 55,940 sentences yourself — the differences between models are far sharper in the source.
• Tom S. Juzek and Zina B. Ward, “Why Does ChatGPT ‘Delve’ So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models,” Proceedings of COLING 2025, 2025. Link ··· A study that couldn’t find the cause in architecture, algorithms, or data, and named human feedback the leading candidate. The experiment design section is the highlight.
• George Orwell, “Politics and the English Language,” Horizon, 1946. Link ··· The original text with the six principles. It’s short — I’d recommend reading it in full.

Background

• Tom S. Juzek et al., “Word Overuse and Alignment in Large Language Models,” arXiv:2508.01930, 2025. Link ··· A follow-up study reconfirming that learning from human feedback is the strongest contributing factor to lexical overuse.

Related issues worth reading together

• Did AI Really Write the Declaration of Independence? ··· If today’s issue is about the player (style), that one is about the referee (the detector). Read them together and you’ll see the whole game.
• The AI Aced the Math Olympiad — With No Proctor in Sight ··· A different branch of the same question: how do you verify what an AI claims?

---

📝 Glossary

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at INLEVEL9. He teaches statistics and data analysis, including business data management and business analytics, while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master's from Korea University's Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Nominalization: the habit of turning what could be a verb into a noun instead — stretching “expand” into “pursue an expansion,” for instance. It makes sentences longer and blurs who is doing what. ↩

2. RLHF (Reinforcement Learning from Human Feedback): a method where human raters choose the better of several answers a model produces, and that preference gets trained into the model. It’s a structure that rewards whatever style catches the rater’s eye. ↩