Humaniser

Writing Guides

Signs of AI writing: common tells and how to fix them

The common signs of AI writing are overused words such as delve and underscore, stock sentence patterns, heavy em dashes, chat-style formatting, an inflated generic tone, and citations that don't check out. Each one is an indicator. Human writers use all of them, so none proves that a machine wrote the text.

By HumaniserPublished Updated 10 min read

We checked every source below on November 10, 2026. The before-and-after examples and the reliability ratings in the table are our own, not quotes or grades from Wikipedia or the studies. The examples use fictional places and people.

Signs of AI writing: the quick list

These nine tells come from Wikipedia's field guide and published studies. Treat any single sign as a reason to read more closely.

Signs of AI writing, how reliable each one is (our reading of the sources), and how to fix it
SignExampleHow reliableHow to fix it
Overused "AI vocabulary"delves, underscores, showcasing, intricate, tapestryWeak alone. Wikipedia calls many in one edit one of the strongest tells. Lists shift as models and writers' habits change.Sources: Wikipedia (WikiProject AI Cleanup), Kobak et al., Science Advances (2025), Juzek and Ward, COLING (2025), Reinhart et al., PNAS (2025), Geng and Trotta, Findings of ACL (2025)Use the plain word you mean: looks at, shows, important.
Inflated significance"...opened in 1962, marking a pivotal moment for the town"Moderate, when a closing -ing phrase adds no fact.Sources: Wikipedia (WikiProject AI Cleanup), Reinhart et al., PNAS (2025)Cut the tail or replace it with the fact it gestures at.
"Serves as" in place of "is""The hall serves as the polling place and boasts a car park."Weak to moderate. Measured across about one million arXiv abstracts, not single sentences.Sources: Wikipedia (WikiProject AI Cleanup), Geng and Trotta, arXiv (2024)Write is or has when that is what you mean.
Negative parallelisms and the rule of three"It's not just a tool, it's a movement." "Fast, flexible and fun."Weak. Wikipedia notes negative parallelism is common in human writing, and the rule of three is a standard device.Source: Wikipedia (WikiProject AI Cleanup)State the point directly. Keep only items you can support.
Vague attributions"Experts argue..." "Observers have noted..."Moderate. Chatbots often present one or two sources as a consensus.Source: Wikipedia (WikiProject AI Cleanup)Name and cite the source, or cut the claim.
Em dashesDashes where a comma or colon would do, often with spaces around themWeak. Wikipedia now calls heavy em dash use a historical indicator, useful only with other signs.Sources: Wikipedia (WikiProject AI Cleanup), Purdue OWLKeep dashes you would choose anyway. Use commas, colons or periods elsewhere.
Chat formatting in finished text** or ## symbols, bold in prose, title-case headings, emoji bulletsStrong for raw Markdown where it doesn't belong. Other quirks matter only in context.Source: Wikipedia (WikiProject AI Cleanup)Format for where the text will appear.
Chatbot leftovers and source disclaimers"Certainly! Here is..." "I hope this helps." "While specific details are limited..."Strong for text addressed to the user. Source disclaimers are among Wikipedia's most common current signs.Source: Wikipedia (WikiProject AI Cleanup)Delete text meant for the asker. Replace a disclaimer with the fact, or cut it.
Citations that don't check outA reference that doesn't exist, a dead link, a DOI that opens an unrelated paperStrong. A fake reference is an error whoever wrote it.Sources: Walters and Wilder, Scientific Reports (2023), Wikipedia (WikiProject AI Cleanup)Open every source and confirm it says what you cite it for.

What does Wikipedia's Signs of AI writing page say?

Wikipedia:Signs of AI writing is an advice page from WikiProject AI Cleanup, first created in December 2023. A banner says it is not a Wikipedia policy. It calls itself a "field guide" to undisclosed AI text on Wikipedia. We read the version current on November 10, 2026. Its own caveats come before its list:

  • Some signs may not apply outside Wikipedia, and fiction tells are left out.
  • Not every text with these signs is AI-generated. The models learned from human writing, and many patterns appear in editorials, blogs and fan fiction.
  • The list is "descriptive, not prescriptive". Readers should not rely only on detectors or their own judgment.
  • "Please do not merely treat these signs as the problems to be fixed." They can point to less visible problems, and those are what need fixing.

As of October 2026, the page names the most common signs of new AI text on Wikipedia as Markdown, AI edit summaries, hedging disclaimers about sources, vague "associated with" statements, and undue emphasis on how records identify a subject.

Which AI words and phrases are overused?

Studies have measured style words, mostly verbs and adjectives, that became far more common after ChatGPT's release:

  • Kobak and colleagues studied more than 15 million PubMed abstracts. In 2024, delves appeared 28 times as often as its 2021 and 2022 trend predicted, underscores 13.8 times and showcasing 10.7 times. They estimate that at least 13.5% of 2024 abstracts were processed with an LLM.
  • Juzek and Ward found 21 "focal words" whose rise in PubMed is likely due to LLM use, including intricate, boasts, realm and groundbreaking.
  • Reinhart and colleagues found that GPT-4o and GPT-4o Mini used camaraderie, tapestry and intricate at more than 100 times the human rate in their comparison corpus.

Lists go stale: Geng and Trotta report a marked drop in delve soon after it was pointed out in early 2024. And these are corpus-level findings. Kobak's team says its analysis "cannot identify individual abstracts" processed by an LLM. One delve in an essay proves nothing.

Before
This report delves into the pivotal role of onboarding, showcasing how a meticulous approach can foster lasting engagement.
After
This report asks whether onboarding affects how long new customers stay. It compares a welcome call with an email-only setup.
Our illustration. Stock words become plain ones, and the sentence says what the report compares.

Which sentence patterns read as AI?

Wikipedia's guide lists these recurring habits:

  • Inflated significance. A closing -ing phrase says why a plain fact matters: "marking a pivotal moment". GPT-4o used present participial clauses at 5.3 times the human rate.
  • "Serves as" for "is". Boasts or features replace has. In arXiv abstracts, is and are fell by more than 10% in 2023 after years of stability.
  • Negative parallelisms. "Not just X, but Y", "It's not X, it's Y" and "Y rather than X".
  • The rule of three. Triplets that make a thin point look complete.

People write all of these. Wikipedia notes that negative parallelism is common in "myths busted" listicles.

Before
The library is not just a building, it's a hub for the community. It opened in 1962, marking a pivotal moment for the town and highlighting its commitment to learning, connection and growth.
After
The library opened in 1962. It runs a homework club on weekday afternoons and lends tools as well as books.
Our illustration. The parallelism, the -ing tail and the triplet give way to two facts.

Is the em dash a sign of AI?

Not on its own. The em dash is standard punctuation, and Purdue OWL teaches it as a way to set off or emphasize content. Wikipedia's guide says chatbots used em dashes more than nonprofessional writers in the same genre, usually with spaces around them, and calls the sign "most useful when taken in combination with other indicators".

The guide now files em dash overuse under historical indicators, dated 2022 to September 2026, and says some AI companies tried to make newer chatbots stop using them. If you have always liked dashes, keep them.

Before
The update is small — but the impact is huge — and users will notice it right away.
After
The update is small, but users will notice it: search results now appear before you finish typing.
Our illustration. A comma and a colon replace the dashes.

Formatting and chatbot leftovers

Wikipedia's guide lists Markdown asterisks and number signs, title-case headings, heavy boldface and emoji on headings or list items. It treats triple-backtick code blocks and citation markup such as contentReference[oaicite:1] as among the clearest signs.

Text meant for the person who asked is another giveaway: "Certainly!" or "I hope this helps". So are source disclaimers such as "based on available information". These signs point to where text may have come from. They say nothing about whether its facts are right.

Before
Certainly! Here is a summary of the meeting. ## Key Takeaways - **Timeline:** Launch moves to May 🚀 - **Ownership:** Priya drafts the support plan by Friday I hope this helps!
After
The team moved the launch to May. Priya will draft the support plan by Friday.
Our illustration. Only the two decisions stay.

Tone: inflated, generic and vaguely sourced

Wikipedia's guide explains the tone problem as regression to the mean. Models favor the common description over the rare, specific fact, so the subject becomes "less specific and more exaggerated". Claims get credited to unnamed experts.

In Russell and colleagues' study, readers who rarely used AI for writing treated any neutral tone as AI, "resulting in many false positives".

Before
Nestled in the heart of the valley, this vibrant town boasts a rich tapestry of traditions. Experts agree it is a must-see destination.
After
The town holds a market every Saturday. Its spring festival dates from the 1890s, according to the town archive.
Our illustration. Checkable details and a named source replace the praise.

Fake or mismatched citations

A citation that doesn't check out is the most serious sign, because it is an error whoever made it. In a 2023 study, Walters and Wilder had ChatGPT write short literature reviews on 42 topics and checked all 636 references: 55% of GPT-3.5 citations and 18% of GPT-4 citations were fabricated. Those models are old now. The check still applies.

Wikipedia's guide adds practical tests: several dead links in a new article, DOIs that don't resolve or lead to unrelated papers, and ISBNs with invalid checksums. It also notes that AI text often has citations now, which says nothing about their accuracy.

Before
Studies show that most readers skim online articles (Smith, 2021).
After
Open Smith (2021) first. If it supports the claim, say what it measured and link it. If you can't find it, cut the sentence.
Our illustration. Here the fix is a check. A fluent sentence with a fake reference is still wrong.

How to tell if something is written by AI, and why tells are not proof

You can't tell for certain from style alone. Studies show why:

  • In Russell and colleagues' study , four readers who rarely used AI for writing spotted AI text at about the rate of random chance on a first batch of 60 nonfiction articles. Five readers who often used AI for writing did far better: across all 300 articles, their majority vote misclassified 1.
  • Liang and colleagues ran seven widely used AI detectors on 91 TOEFL essays by non-native English writers. The detectors labeled more than half of them AI-generated, while scoring essays by US eighth graders almost perfectly.

Wikipedia's guide warns that false accusations of AI use can drive away new editors. It lists perfect grammar, "fancy" prose and transition words on their own as ineffective indicators. One sign of human writing it names is a writer who can explain their own editorial choices.

If you are judging someone else's work, treat tells as a reason to ask for drafts or version history and to talk to the writer. Our guide to AI detector false positives covers how writers can show their process, and how AI detectors work explains why scores vary.

Checklist: how to make your writing not sound like AI

Fix the substance first: Wikipedia's editors warn that treating the signs themselves as the problem could just make detection harder. If you used AI, follow the disclosure rules that apply to you, such as your course or publisher policy. Our responsible AI use guide covers what to settle first.

  1. Check every fact, number and name against a source you have opened.
  2. Open every citation and confirm it supports its sentence.
  3. Replace general praise with a specific detail.
  4. Name who said what, or cut "experts say".
  5. Delete closing -ing phrases that add no fact.
  6. Swap stock words for plain ones, unless the stock word is accurate.
  7. Write is and has where they fit.
  8. Rewrite "not X, but Y" as one direct claim.
  9. Remove chat greetings, Markdown symbols and emoji the format doesn't call for.
  10. Read it aloud and change anything you wouldn't say.

For a method that keeps every claim intact while you edit, see how to humanize AI text without changing meaning.

Check a draft

Humaniser's free AI detector gives 30 checks a day with no account. It is a free first check: scores can differ between detectors, and no score proves who wrote a text.

The Humaniser humanizer rewrites a passage you paste in. The free plan allows 3 humanizations a day of up to 200 English words each. Compare every change with your original, and use it only where your course or publisher allows that kind of help.

Common questions

Is there an official list of AI words to avoid?

No. Lists come from different studies of different texts, and they change as models and writers' habits change. Wikipedia sorts its list by model era and notes that an overused word does not make its synonyms suspect. Use the word that is accurate.Sources: Wikipedia (WikiProject AI Cleanup), Kobak et al., Science Advances (2025), Geng and Trotta, Findings of ACL (2025)

Can a teacher prove AI use from writing style alone?

No. Kobak and colleagues say their word-frequency method cannot identify individual abstracts, and in another study readers who rarely used AI for writing spotted AI text at about the rate of chance. Drafts, version history and a conversation with the writer are better evidence.Sources: Kobak et al., Science Advances (2025), Russell, Karpinska and Iyyer, ACL (2025), Wikipedia (WikiProject AI Cleanup)

What are the signs of AI writing in fiction?

Wikipedia's guide covers informational writing and leaves fiction tells out. In a 2025 PNAS study, GPT-4o and GPT-4o Mini used words such as camaraderie and tapestry at more than 100 times the human rate across six genres including fiction, as in the GPT-4o line "The camaraderie was palpable." Novelists use these words too.Sources: Wikipedia (WikiProject AI Cleanup), Reinhart et al., PNAS (2025)

Do AI detectors look for these signs?

Detectors use their own models, and scores can differ between tools. In a 2023 study, the non-native English essays flagged by all seven detectors had significantly lower perplexity, meaning more predictable wording. The authors suggest detectors may penalize non-native writers with limited linguistic expressions.Source: Liang et al., Patterns (2023)

Sources

Checked on November 10, 2026. Wikipedia pages change often; the Wikipedia section above describes this revision.