(Warning: Rant ahead. Click here to get to the meat of this post.)

Every time I see an em dash, I cringe. And I wince. And my hackles go up.

In other words, it bugs the hell out of me.

This innocuous little punctuation mark didn’t bother me in the slightest until generative AI tools started including the dang thing EVERYWHERE. Now I’m paranoid about using it in my own writing because it’s one of the obvious signs that AI led the way in that piece.

No AI Shaming Here

Don’t get me wrong. I love it when people make use of this incredible resource at everyone’s fingers. I mean, if AI can write a 14-page research report in three minutes, why wouldn’t we use that as a first draft instead of spending three weeks trying to do it ourselves?

AI Writing Should Be a First Draft

My problem comes when we use AI to generate copy that isn’t true to our authentic voice, especially when we’re trying to convey authentic feelings or sincere points.

Maybe it’s just the writer in me. I have a master’s in journalism, and writing comes as naturally as breathing to me (most of the time). So when I see LinkedIn posts replete with the signs of AI writing, I’m, well, judgy.

I shouldn’t feel this way. Ever since I discovered generative AI (almost two years before ChatGPT showed up), I’ve called it the great equalizer. I recognize its ability to help people express themselves, no matter their writing level. This is a wonderful thing. But now that everyone has this ability, it’s my humble opinion that we owe it to our readers to add our own voices to the content we share.

Pot, Meet Kettle

Hi, everyone. My name is Beth, and I’m a big hypocrite.

As a speaker, I need to have a super-duper online presence on social media. And I wasn’t keeping up, so this year I hired a social media team to share some of my content. And I created a custom GPT called BethBot that helps the team capture more of my voice.

Even though I specifically told the BethBot to avoid AI writing flags, it still uses em dashes and other telltale signs. And every time I see content with my name on it with those flags, I wince/cringe/sigh and die a little inside. I’m still trying to figure out if the discomfort outweighs the benefit.

Rant over. Now on to the point of this post.

Signs that Content Was Written by a Bot

AI-generated content uses particular word patterns, phrases, punctuation and other signs that will help you identify its bot origins. Scientists have analyzed the content to find the patterns, and I’m going to share some big AI-content indicators here.

Here’s a link to a conversation with ChatGPT where I asked it to read that study and create a list. I also put the info into an infographic below.

And here are my top 20 ick phrases, not including the dreaded em dash.

1. Delve into​
2. In the fast-paced world of…​
3. Harnessing the power of​
4. It is worth noting that​
5. Plays a pivotal role in​
6. A plethora of​
7. In today’s digital landscape
8. Seamless integration of​
9. Catalyst for change​
10. Amidst the evolving…​
11. Paves the way for​
12. At the forefront of​
13. Unveiling the potential of​
14. A cornerstone of​
15. Revolutionizing the way we​
16. A game-changer in​
17. Bridging the gap between​
18. From inception to execution​
19. The ever-expanding realm of​
20. Underscores the importance of

Download this image and share!

More Insights from the Study

I asked ChatGPT to read that paper and extract some key findings. It’s fascinating:

Characteristic Phrases & Word Choices

Each LLM tends to favor certain phrases. The study found that:

• ChatGPT prefers “such as,” “certainly,” “overall,” “sure,” “utilize,” “various,” “typically.”
• Claude favors “here,” “according to,” “based on,” “the text,” “while,” “appears to,” “both,” “when.”
• Grok frequently uses “which,” “where,” “not,” “here,” “here is,” “remember.”
• Gemini often has “might,” “but also,” “not only,” “helps in.”
• DeepSeek likes “below is,” “in summary,” “for example,” “may,” “certainly! below,” “know if,” “example.”

These words and phrases could serve as red flags when identifying text generated by a particular LLM.

First-Word Preferences

The study also highlights that the first word of an LLM-generated response can be a telltale sign:

• ChatGPT: “as,” “yes,” “sure,” “here,” “in,” “to,” “creating,” “certainly,” “title,” “the.”
• Claude: “in,” “from,” “this,” “how,” “yes,” “title,” “according,” “the,” “based,” “here.”
• Grok: “step,” “introduction,” “yes,” “creating,” “to,” “title,” “in,” “certainly,” “the,” “here.”
• Gemini: “my,” “creating,” “rain,” “while,” “here,” “yes,” “elara,” “python,” “this,” “the.”
• DeepSeek: “based,” “yes,” “step,” “comprehensive,” “here,” “to,” “creating,” “title,” “certainly,” “the.”

Markdown Formatting Idiosyncrasies

• ChatGPT often uses bolded text for key points, structured headings, and detailed markdown formatting with indentation and lists.
• Claude prefers minimal formatting, using plain lists or numbered lists with basic structure.
• Gemini frequently uses italics more than others.
• DeepSeek leans towards “here’s a breakdown” and bullet lists.

Sentence Starters & Transitional Elements

Certain transitions and sentence structures are associated with specific models:

• ChatGPT: Uses “Certainly,” “In summary,” “Such as,” “Overall.”
• Claude: References the prompt often with “According to the text,” “Based on,” “Here’s a summary.”
• DeepSeek: Uses “Below is,” “Here’s a breakdown,” “Key improvements.”

Punctuation & Text Structure Patterns

• ChatGPT tends to use structured bullet points, subheadings, and bold emphasis.
• Claude minimizes markdown but uses clean, readable formatting without excessive styling.
• Grok & DeepSeek sometimes lean on structured responses but lack ChatGPT’s emphasis formatting.
• Gemini has more frequent use of italics compared to others.

These findings suggest that certain words, phrases, punctuation styles, and markdown habits are unique fingerprints of each LLM. If you’re looking to filter or identify LLM-generated content, focusing on sentence starters, characteristic transitions, markdown patterns, and first-word choices can be a strong approach.