AI slop
Signs of AI writing: how to spot it in posts, replies and emails
The signs of AI writing that research and Wikipedia’s editors agree on, which ones no longer work, and how well people really spot AI text, with examples.
Short answer: The signs that hold up are words chatbots overuse, sentence shapes like “It’s not just X, it’s Y” and lists of three, puffed-up importance such as “a testament to”, chatbot openers and sign offs, a customer service tone, and replies that could sit under any post. Perfect grammar and a formal tone aren’t signs, and the em dash, the most famous one, is fading as new models drop it. One sign proves little; several together make a strong case. People who use ChatGPT a lot are very good at this: in a 2025 study, five of them voting together got 299 of 300 texts right.
The signs that hold up
The fullest list comes from Wikipedia’s editors, who have kept a guide to AI writing since December 2023, because they clean up articles written by chatbots. They call it a field guide, and say so plainly:
this list is descriptive, not prescriptive; it consists of observations, not rules
- 1A chatbot opener
- 2Puffed-up importance
- 3A trailing “-ing” phrase
- 4The fake contrast
- 5A chatbot sign off
| Sign | What it looks like | What a person would write |
|---|---|---|
| Chatbot words | delve, tapestry, testament, pivotal, crucial, showcase | Look into, shows, key, show |
| The fake contrast | It’s not just a tool, it’s a movement. | Say what the tool does. |
| Three in a row | Fast, simple, and powerful. | Keep the one that’s true. |
| A trailing “-ing” | The team shipped it, highlighting their commitment to users. | The team shipped it. |
| Puffed-up importance | a pivotal moment, an evolving landscape, a lasting mark | Say what changed. |
| Chatbot manners | Great question! I hope this helps. Let me know if you have questions. | Start with the answer, and stop. |
| Customer service tone | I understand your concerns about this. | Answer the concern. |
| Hedging | While specific details are limited, based on available information... | Say what you know. |
Research backs most of these. A study in PNAS found that instruction-tuned models use “-ing” clauses two to five times as often as people do, and that GPT-4o uses words like “tapestry” and “camaraderie” about 150 times as often. Source: Reinhart and others, 2025.
The Washington Post’s look at 328,744 ChatGPT messages found versions of “not X, but Y” in 6% of chats by July 2025, and an emoji in 70% of messages. Source: The Washington Post, via The Boston Globe, November 13, 2025.
The signs in replies and posts
Under a post on X, Threads or Bluesky, the tells get shorter, and one stands out from the rest: the reply could sit under any post. “Great insights! Thanks for sharing.” says nothing about the post it answers. We show it under three posts that have nothing in common in what AI slop is.
Other short tells:
- Praise up front, then a slogan: “Love this perspective. Consistency is key.”
- A question about nothing in the post: “What’s your take on the future of AI?”
- Emoji used as bullet points in a reply that didn’t need a list.
You can paste any reply into our free AI Slop Checker to see which of these lines it has, with what to write instead.
Signs that don’t work, or no longer do
Wikipedia’s editors keep a list of signs people trust that don’t hold up:
- Perfect grammar. Many people write cleanly.
- A formal tone, or prose that feels “bland” or “robotic”.
- Transition words. Only a few are overused by chatbots.
And a list of signs that used to work:
- The em dash. Wikipedia now files it as a sign of AI writing from 2022 to September 2026, because newer models, such as OpenAI’s GPT-5.1, were made to avoid it. Our article on why AI uses so many em dashes tells the whole story.
- Summary endings like “In summary” or “In conclusion”, which older models added to everything.
- Lines about a knowledge cutoff, from the first chatbots.
How well people spot AI writing
Most people guess, and guess badly:
- About 50%. In a 2023 study of 4,600 people, readers spotted AI-written profiles no better than a coin toss. They trusted the wrong signs, such as “I”, contractions and talk about family, as proof a person wrote it. Source: Jakesch, Hancock and Naaman, PNAS 2023.
- Close to chance for people who rarely use chatbots, in a 2025 study.
- 299 of 300 for five people who use ChatGPT for writing often, voting together, in that same study. Their most common clue was vocabulary, in over half their explanations, then sentence structure such as lists of three.
the majority vote among five such “expert” annotators misclassifies only 1 of 300 articles
The more AI writing people read, the faster they spot it. Your readers on X read a lot of it.
What to do with a sign
- Don’t accuse anyone on one sign. People wrote “pivotal” long before chatbots did. Several signs together say much more.
- Check your own writing. If you draft with AI, cut the chatbot words, the fake contrasts and the openers, and add one thing only you could say.
- In replies, mention the post. A number from it, a word the author used, what’s in the picture. No chatbot template can fake that.
Twick drafts replies with 100% no AI slop, each about the post in front of you and in your own voice. You pick one and post it yourself.
Questions
What are the signs of AI writing?
Chatbot words such as delve and tapestry, fake contrasts, lists of three, trailing “-ing” phrases, puffed-up importance, chatbot openers and sign offs, a customer service tone, and replies that could sit under any post. Several together are a strong sign.
Can you tell if a text was written by AI?
Often, if you read a lot of AI writing. In a 2025 study, people who use ChatGPT often were right almost every time, while people who rarely use it did about as well as chance.
Are em dashes a sign of AI?
Less every month. They were one of the best known signs from 2022, but newer models avoid them, and Wikipedia now lists them as a sign that used to work.
Is perfect grammar a sign of AI?
No. Wikipedia’s guide lists perfect grammar and a formal tone among the signs that don’t work.
What words does AI use the most?
It changes with each model. The best known are delve, tapestry, testament, pivotal and underscore. Our list of AI words and phrases has the rest, with what to write instead.