Est.

Signs of AI Writing: A Field Guide for Editors

The patterns that give machine-written prose away, the ones that stopped working, and a short check you can run on any draft before it ships.

High Quality Writing EditorsUpdated

Most signs of AI writing are habits, not errors. A language model predicts the likeliest next word, then gets tuned toward answers the widest range of people find agreeable. The output drifts toward the average: the specific, odd, checkable detail gets swapped for a general, positive one. That smoothing is what an attentive reader picks up on, and it shows in the content, the word choice, the formatting and the citations.

No single sign below proves anything. People wrote every one of these patterns before chatbots existed, and models learned them from people. What gives text away is several signs at once, and the absence of anything only its author could have known.

Read for clusters, not single words

One "pivotal" in a paragraph is a word choice. "Pivotal", "underscores", "fostering" and "rich tapestry" in the same paragraph, wrapped around a claim with no names, numbers or dates in it, is a pattern. Score a passage on how many independent signs it shows, from different families below, rather than on the strongest one.

The patterns also move. The words, punctuation and structures models overuse change from one model generation to the next, and the most notorious tells are the first ones vendors train out. A list that worked in 2024 will flag the wrong things in 2026.

Signs in what the text says

Inflated significance

Ordinary facts get promoted to turning points. Watch for "marks a pivotal moment", "stands as a testament to", "plays a crucial role", "reflects broader trends", "setting the stage for" and "the evolving landscape". The giveaway is that the sentence would fit any subject: swap the noun and it still reads.

Analysis tacked on with an -ing clause

A plain fact, then a comma and a present participle that tells you what it means: "..., highlighting its importance", "..., ensuring a seamless experience", "..., underscoring the need for". The clause adds an interpretation nobody in the text has argued for.

Vague authorities

"Experts argue", "observers have noted", "industry reports suggest", "several sources" when one is cited, or no source at all. Human writers do this too; models do it reflexively, and newer ones with web search will pin the opinion on a named source that never said it.

Generic where a fact should be

The detail a person would remember (the year, the name, the number, the odd exception) is missing, and a positive generality sits in its place. "A key player in the industry" instead of what the company actually makes. This is the strongest content sign, because it is also what makes the text worse.

The formula ending

A closing section on "challenges and future outlook" that opens with "Despite these challenges" and ends on vague optimism, or a final paragraph that restates the piece and starts with "Ultimately" or "In conclusion".

Signs in the language

Vocabulary clusters

Researchers have measured a jump in specific words in text written after chatbots became widely available, "delve" being the best-known example. Wikipedia's editors keep a running list of these words, grouped by the model era they belong to. A selection from their list:

Roughly whenWords the list flags
2023 to mid-2024additionally, boasts, crucial, delve, intricate, landscape, meticulous, pivotal, tapestry, testament, underscore, vibrant
Mid-2024 to mid-2025align with, bolstered, enhance, enduring, fostering, highlighting, showcasing
Mid-2025 to mid-2026emphasizing, enhance, highlighting, showcasing
Mid-2026 ondependable, echoed, leverage, matters, multifaceted, practical, prioritize, quietly, steady

Read the list literally: an overused word does not make its synonyms suspicious, and the same word in its ordinary sense (an underscore character, a landscape painting) means nothing. Check the current list on Wikipedia before relying on it; it is revised as models change.

Avoiding "is" and "has"

Models reach past the plain verb: "serves as", "stands as", "functions as", "represents" instead of "is"; "boasts", "features", "offers" instead of "has". People write "the gallery is" and "the city has" far more than a model does.

"Not just X, but Y"

Negative parallelism corrects a misconception nobody raised: "It's not just a tool, it's a philosophy." "This isn't about speed. It's about trust." One is a flourish; one per section is a template.

Everything in threes

Three adjectives, three short phrases, three bullet points, again and again, often in places where a person would not bother with a flourish at all. The triplet makes a thin point look complete.

Even rhythm

Sentences of similar length and shape, paragraph after paragraph. Human prose speeds up and slows down; a short sentence lands after a long one. Detection tools measure this as burstiness, and you can hear it by reading a paragraph aloud. We cover the mechanics in Perplexity and Burstiness as Detection Signals and why training produces it in How Training Objectives Shape the Lifeless Rhythm of ChatGPT Prose.

Signs in the formatting

  • Title Case In Every Heading, where the publication otherwise uses sentence case.
  • Bold scattered through running prose, or stray asterisks where bold didn't render.
  • Bullet lists where each item opens with a bolded label and a colon, then a sentence.
  • Emoji at the start of headings or list items.
  • Markdown left in places that don't render it: ### before a heading, ** around words, --- as a divider.
  • Curly quotes and apostrophes pasted into a system that otherwise uses straight ones.
  • Chatbot citation debris such as contentReference[oaicite:1] left in the text.

Citation debris is close to proof. The rest are house-style questions that only become signs when they cluster.

Leftovers from the chat window

The plainest signs are lines that were written to the person prompting, not to the reader, and survived the copy and paste:

  • Openers and closers: "Certainly! Here's a...", "Great question", "I hope this helps", "Let me know if you'd like me to expand on any section."
  • Template slots never filled in: a name, date or company left in square brackets for the writer to replace, or links to example.com.
  • Disclaimers about the model's own sources: "based on available information", "specific details are limited", "as of my last update".
  • Advice to the writer, not the reader: "verify these figures before publishing".

Signs in the citations

Models now cite sources more often than they used to, so the presence of citations proves nothing. Check them:

  • Links that return a 404 and have no copy in the Internet Archive, which suggests they never existed.
  • DOIs that don't resolve, or resolve to an unrelated paper; ISBNs with an invalid checksum.
  • Real sources that don't say what the sentence attributes to them. This is the slowest one to catch: open the source and find the claim.
  • Book citations with no page number, and tracking parameters such as utm_source=chatgpt.com left on URLs.

Signs that don't hold up

False accusations are costly, and several popular tells are weak or have stopped working:

  • The em dash. Heavy, formulaic em dash use was a real pattern in chatbot output for years. It became so well known that vendors tuned newer models to suppress it, and Wikipedia's editors now list it as a historical indicator rather than a current one. Plenty of professional writers use em dashes; one tells you nothing.
  • Perfect grammar. Many people write cleanly, and many edit with tools that fix grammar without writing anything.
  • Formal or academic prose. Models overuse particular words, not formality in general.
  • Transition words on their own. "Additionally" and "Moreover" are old essay habits as much as model habits.
  • Prose that feels bland. Blandness is a quality problem whoever wrote it, and some model output doesn't read as bland at all.

Signs that point the other way: text that demonstrably predates November 30, 2022 (when ChatGPT was released to the public), and a writer who can explain, without notes, why they made a specific choice in the piece.

What about AI detectors?

Detectors do better than chance and still make enough mistakes, in both directions, that a score shouldn't decide anything on its own. They can be thrown off by paraphrasing, light editing and models they weren't trained on, and they flag some human writing as AI. Use a score as one more signal next to the ones above. We look at the evidence in How Accurate Are AI Detectors and at the human-written text they misflag in AI Content Detectors Rated for False Positive Risk.

A five-minute check for any draft

  1. Search the text for chatbot leftovers and placeholders: "Certainly", "I hope", "Let me know", square brackets, "oaicite", "utm_source".
  2. Open every link and citation. Confirm each one exists and says what the sentence claims.
  3. Pick the densest paragraph and count signs from different families: inflated significance, an -ing tail, a vocabulary cluster, a "not just X", a triplet.
  4. Find one sentence only this author could have written: a name, a number, a date, a first-hand detail. If there isn't one, that matters more than any word list.
  5. Read two paragraphs aloud. Listen for sentences that all land at the same length.
  6. If it still matters who wrote it, ask the writer about one specific choice in the piece.

If you're the one using AI to write

Deleting every "delve" doesn't fix a draft. The signs are symptoms: the generic claim, the missing fact and the unsupported interpretation are the problems, and a reader notices them whether or not they think about AI. Fix those and most of the surface signs go with them.

Teams that publish at volume turn this list into checks that run on every draft, the way code runs tests. We cover that in Writing Test Cases as Quality Gates and Filler Pattern Detection in AI-Generated Drafts, and compare rewriting a finished draft with changing how it's generated in Comparing Humanization Tools to Adversarial Writing Systems.

Frequently asked questions

How can you tell if something was written by AI?

Look for several signs at once: inflated claims about significance, analysis tacked on with -ing clauses, clusters of words models overuse, "not just X, but Y" constructions, even sentence rhythm, leftover chatbot phrases and citations that don't check out. Above all, look for missing specifics. No single sign is proof.

Is an em dash a sign of AI writing?

Not on its own. Formulaic em dash use was common in chatbot output, but newer models have been tuned to avoid it and many human writers use em dashes. Treat it as weak evidence at most, and only alongside other signs.

What words does AI overuse?

It depends on the model and the year. Earlier chatbot text leaned on words like delve, tapestry, testament, pivotal and underscore; later models favored enhance, fostering, highlighting and showcasing. Wikipedia's "Signs of AI writing" page keeps a current list by model era.

Can AI detectors prove text was written by AI?

No. Detectors beat chance but make errors in both directions, and editing or paraphrasing changes their scores. Use a detector result as one signal, never as the deciding one.

Sources

  1. Wikipedia:Signs of AI writing (WikiProject AI Cleanup)
  2. Delving into LLM-assisted writing in biomedical publications through excess vocabulary (Kobak et al.)

Go deeper