---
title: Why AI-Written Text Is So Easy to Spot | Nobu Sato
description: Five frequent ChatGPT users misjudged 1 of 300 articles. 379 of 454 words that surged in 2024 abstracts were style words. What gives AI prose away, and why.
image: https://nobusato.jp/hubfs/covers-2026-11/why-ai-writing-is-easy-to-spot-cover-en-2026-10-m.png
---

[Nobu Sato](https://nobusato.jp/en?hsLang=en)

[Profile](https://nobusato.jp/en/profile?hsLang=en) [Method](https://nobusato.jp/en/blog/method?hsLang=en) [Articles](https://nobusato.jp/en/blog?hsLang=en) [Glossary](https://nobusato.jp/en/blog/glossary?hsLang=en) [Contact](https://nobusato.jp/en/contact?hsLang=en)

[日本語](https://nobusato.jp/ja/blog/why-ai-writing-is-easy-to-spot)[English](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot)[中文](https://nobusato.jp/zh/blog/why-ai-writing-is-easy-to-spot)

Career

In this article

1. [Can people really tell?](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s1)
2. [The first giveaway is vocabulary](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s2)
3. [What English AI prose sounds like](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s3)
4. [Why does it write like that?](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s4)
5. [Fixing the words isn't enough](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s5)

1. [Home](https://nobusato.jp/en?hsLang=en)
2. [Articles](https://nobusato.jp/en/blog?hsLang=en)
3. [Career](https://nobusato.jp/en/blog/tag/career)

# Why AI-Written Text Is So Easy to Spot

Words, templates and specifics

[Nobu Sato](https://nobusato.jp/en/profile?hsLang=en)7 min read

Contents5

1. [Can people really tell?](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s1)
2. [The first giveaway is vocabulary](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s2)
3. [What English AI prose sounds like](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s3)
4. [Why does it write like that?](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s4)
5. [Fixing the words isn't enough](https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot#s5)

People who use ChatGPT a lot can spot AI-written text almost perfectly. In one study, five of them voting together misjudged 1 article out of 300, and beat most commercial detectors. So recognizing it is the easy part. The harder question is why it all comes out sounding the same, and what that says about your own writing.

## Can people really tell?

Yes, if they use these tools themselves. Jenna Russell, Marzena Karpinska and Mohit Iyyer presented the study at ACL 2025. They took 300 English non-fiction articles, some written by people and some generated by GPT-4o, Claude and o1. Five annotators who regularly use language models for writing judged each one. Their majority vote got 299 right.

That vote outperformed most of the commercial and open-source detectors they tested, and it held up when the AI text had been paraphrased or run through "humanizer" tools.

What were they looking at? Mostly words, according to the paper, which calls them "AI vocabulary." Beyond that, harder things to measure: text that was too formal, had nothing original in it, and was clearer than it had any reason to be.

The detectors fail in the other direction. In 2023, Weixin Liang, James Zou and colleagues at Stanford tested seven widely used detectors on 91 TOEFL essays written by people. On average, the detectors flagged 61.22% of them as AI. Eighty-nine of the 91 were flagged by at least one detector. Essays by US eighth graders were classified almost perfectly.

Non-native writers use more common words and more regular sentence patterns, and that is what the detectors read as machine output. If you work in Japan and write English as a second language, your plain, careful emails are the ones most likely to be called fake.

It cuts the other way for employers too. If your hiring team in Japan runs candidates' English writing samples through a detector, the 2023 study suggests the people most at risk of a false flag are careful non-native writers, which describes most of your bilingual applicants. I would not let a detector score anywhere near a hiring decision.

## The first giveaway is vocabulary

Here the numbers are unusually solid. Dmitry Kobak and colleagues looked at more than 15 million biomedical abstracts indexed in PubMed from 2010 to 2024 and measured which words suddenly became more common. They estimate that at least 13.5% of 2024 abstracts were processed with a language model, reaching 40% in some subsets by field, country and journal. The paper appeared in Science Advances in 2025.

The kind of word is what stands out. In earlier years, the words that spiked were mostly nouns tied to what researchers were studying, like "masks" during the pandemic. In 2024, 379 of the 454 excess words were style words, and two thirds of those were verbs. "Delves" appeared at 28.0 times the frequency the earlier trend predicted. "Underscores" was at 13.8 times, "showcasing" at 10.7 times.

Most of the words that surged in 2024 research abstracts were style words

Basis: over 15 million biomedical abstracts in PubMed, 2010 to 2024. Ratio = actual 2024 frequency divided by the frequency expected from the prior trend. Source: Kobak et al. (2025), Science Advances

**Excess ratio**　less common words

- delves

  28.0×
- underscores

  13.8×
- showcasing

  10.7×

**The 454 excess words of 2024**　by type

- Style wordstwo thirds are verbs

  379
- Content wordstied to the research topic

  75

Vermilion = style words, unrelated to what the paper is about

What grew was the phrasing. The content stayed the same size.

Figure 1. Kobak et al. estimate from over 15 million PubMed abstracts that at least 13.5% of 2024 abstracts were processed with a language model; 379 of that year's 454 excess words were style words, and "delves" ran at 28.0 times its expected frequency.

The research didn't change. A coat of paint went on top, and readers can smell the paint.

## What English AI prose sounds like

Beyond single words, there are habits of structure. These are my own observations, not findings from a study.

It opens with scenery. "In today's fast-paced business landscape." The reader came for a point and gets a weather report.

It leans on signposts. "It's worth noting." "Ultimately." "In conclusion." A real turn in an argument rarely needs a label.

It loves the setup and reversal: "It's not just a tool, it's a mindset." Used once, nobody notices. The trouble is that it comes back in paragraph four, and again in paragraph seven. The em dash belongs to the same family, the dramatic pause before the reveal.

It groups things in threes, and every paragraph ends on a line built to be quoted.

Japanese and Chinese have their own versions. Japanese AI prose stacks polite hedges like 「〜と言えるでしょう」 (one could say) at the end of every paragraph and closes with a section literally titled まとめ, summary. Chinese AI prose opens with 「在……的浪潮下」 (amid the wave of) and keeps reaching for 「值得注意的是」 (it is worth noting).

Any of these, on its own, is something people do. What gives AI away is that they all show up at once, spread evenly. Paragraphs of the same length, sections with the same shape, as if cast from one mold.

## Why does it write like that?

As I understand it, there are two layers to the answer.

The first follows from how the models work. A language model predicts the most likely next word. The most likely word is the one everyone uses. The most likely structure is the one taught in writing class. Average everyone's style together and you get a writer who sounds like no one in particular.

The second is the finishing stage. Before release, models go through reinforcement learning from human feedback (RLHF) to make answers match what people prefer. Robert Kirk and colleagues reported at ICLR 2024 that RLHF significantly reduces output diversity compared with supervised fine-tuning alone. The answers get more reliable and more alike.

Past that point I'm guessing. People rating answers probably reward ones that are well organized and neatly summarized. Amplify that preference over many rounds and you get prose that always wraps up.

## Fixing the words isn't enough

Delete every "delve" and the text still reads as AI. Vocabulary is only the surface.

What's really missing is specifics. No named person, no year, no company, no price. No detail only the writer could know. No doubt either: every sentence is equally sure of itself. When I wrote about [AI and interview notes](https://nobusato.jp/en/blog/ai-interview-notes-limits?hsLang=en), the problem was the same. The summary is so smooth that "we did it" turns into "he led it" without anyone noticing.

So what AI flattened for writers is the expression layer. What still has value is what you worked out yourself, including the calls you got wrong. I made that argument about skills in [After AI, which layer of you loses value first?](https://nobusato.jp/en/blog/which-layer-loses-value-first?hsLang=en), and I think it applies to writing just as well.

The test I'd use is narrower than any detector. Look for one sentence in the draft that only the writer could have produced: a number they worked out themselves, or a call they got wrong. If there isn't one, it hardly matters who or what wrote the rest.

Sources (as of November 2026)

- [Jenna Russell, Marzena Karpinska, Mohit Iyyer, "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text," ACL 2025](https://aclanthology.org/2025.acl-long.267/)
- [Dmitry Kobak, Rita González-Márquez, Emőke-Ágnes Horvát, Jan Lause, "Delving into LLM-assisted writing in biomedical publications through excess vocabulary," Science Advances 11(27), eadt3813, July 2025](https://doi.org/10.1126/sciadv.adt3813)
- [Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, James Zou, "GPT detectors are biased against non-native English writers," Patterns, 2023](https://arxiv.org/abs/2304.02819)
- [Robert Kirk et al., "Understanding the Effects of RLHF on LLM Generalisation and Diversity," ICLR 2024](https://proceedings.iclr.cc/paper_files/paper/2024/hash/5a68d05006d5b05dd9463dd9c0219db0-Abstract-Conference.html)

---

Author

[Nobu Sato](https://nobusato.jp/en/profile?hsLang=en)

Launching in Japan, starting with people and systems. I write about building organizations, sales teams and technology organizations in Japan, in English, Japanese and Chinese.

[Profile](https://nobusato.jp/en/profile?hsLang=en)[Contact](https://nobusato.jp/en/contact?hsLang=en)

## Read next

1. 01[Before you change jobs, do a four-layer self-portrait](https://nobusato.jp/en/blog/four-layer-self-portrait-before-changing-jobs?hsLang=en)4 min
   
   A recruiter offers 20% more. Before touching your resume, line up the real reasons behind your last three moves. What ...
2. 02[Pick your next company with a market map](https://nobusato.jp/en/blog/pick-your-next-company-with-a-market-map?hsLang=en)6 min
   
   In Japan, the employer most likely to pay you well hires your function cluster and competes with your current employer. ...
3. 03[Which circle of the talent map are you in?](https://nobusato.jp/en/blog/which-circle-of-the-talent-map-you-are-in?hsLang=en)4 min
   
   Whether recruiters call has little to do with your resume. It is where you sit on the map: same field, adjacent, or ...

[Column guide: Career](https://nobusato.jp/en/blog/guide-career?hsLang=en) · [Glossary](https://nobusato.jp/en/blog/glossary?hsLang=en) · [The method behind the articles](https://nobusato.jp/en/blog/method?hsLang=en)

[← All articles](https://nobusato.jp/en/blog?hsLang=en)

## Writing about launching in Japan, in three languages.

[Profile](https://nobusato.jp/en/profile?hsLang=en) [Get in touch](https://nobusato.jp/en/contact?hsLang=en)

Nobu Sato

Writing about launching in Japan, in three languages.

[info@nobusato.jp](mailto:info@nobusato.jp) [Substack](https://nobusato.substack.com)[Newsletter](https://www.linkedin.com/newsletters/the-japan-entry-memo-7519750783555518464/)[Medium](https://medium.com/@nobusato5888)[LinkedIn](https://www.linkedin.com/in/realnobusato/)[X](https://x.com/nobusato5888) [RSS](https://nobusato.jp/en/blog/rss.xml?hsLang=en)

[Profile](https://nobusato.jp/en/profile?hsLang=en)[Method](https://nobusato.jp/en/blog/method?hsLang=en)[Articles](https://nobusato.jp/en/blog?hsLang=en)[Glossary](https://nobusato.jp/en/blog/glossary?hsLang=en)[Contact](https://nobusato.jp/en/contact?hsLang=en)

Views here are my own and don’t represent any organization I work or have worked for.

© 2026 Nobu Sato. All rights reserved.·[Privacy and disclaimer](https://nobusato.jp/en/privacy?hsLang=en)

```json
{
  "@context" : "https://schema.org",
  "@type" : "BreadcrumbList",
  "itemListElement" : [ {
    "@type" : "ListItem",
    "item" : "https://nobusato.jp/en",
    "name" : "Home",
    "position" : 1
  }, {
    "@type" : "ListItem",
    "item" : "https://nobusato.jp/en/blog",
    "name" : "Articles",
    "position" : 2
  }, {
    "@type" : "ListItem",
    "item" : "https://nobusato.jp/en/blog/tag/career",
    "name" : "Career",
    "position" : 3
  }, {
    "@type" : "ListItem",
    "item" : "https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot",
    "name" : "Why AI-Written Text Is So Easy to Spot",
    "position" : 4
  } ]
}
```

```json
{
  "@context" : "https://schema.org",
  "@id" : "https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot",
  "@type" : "WebPage",
  "about" : {
    "@type" : "Thing",
    "name" : "Career"
  },
  "author" : {
    "@id" : "https://nobusato.jp/#person"
  },
  "description" : "Five frequent ChatGPT users misjudged 1 of 300 articles. 379 of 454 words that surged in 2024 abstracts were style words. What gives AI prose away, and why.",
  "inLanguage" : "en",
  "isPartOf" : {
    "@id" : "https://nobusato.jp/en/blog",
    "@type" : "Blog",
    "name" : "Articles"
  },
  "keywords" : "Career",
  "name" : "Why AI-Written Text Is So Easy to Spot",
  "url" : "https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot"
}
```

```json
{
  "@context" : "https://schema.org",
  "@id" : "https://nobusato.jp/#person",
  "@type" : "Person",
  "alternateName" : [ "佐藤伸明", "Nobu Sato", "左滕" ],
  "description" : "Japan market launches: organization, sales and technology",
  "knowsAbout" : [ "Japan market entry", "Organization design", "Talent acquisition", "Sales organization building", "Technology organization building", "Japanese labor law" ],
  "name" : "Nobu Sato",
  "sameAs" : [ "https://www.linkedin.com/in/realnobusato/", "https://x.com/nobusato5888" ],
  "url" : "https://nobusato.jp/en/profile"
}
```

```json
{
  "@context" : "https://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Nobu Sato",
    "url" : "https://nobusato.jp/en/blog/author/nobu-sato"
  },
  "dateModified" : "2026-11-04T17:40:38.081Z",
  "datePublished" : "2026-11-04T17:46:02.000Z",
  "headline" : "Why AI-Written Text Is So Easy to Spot",
  "image" : [ "https://nobusato.jp/hubfs/covers-2026-11/why-ai-writing-is-easy-to-spot-cover-en-2026-10-m.png" ],
  "mainEntityOfPage" : {
    "@id" : "https://nobusato.jp/en/blog/why-ai-writing-is-easy-to-spot",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "url" : "https://nobusato.jp/hubfs/seo/logo-nobusato-2026-10.png"
    },
    "name" : "nobusato.jp"
  }
}
```