Can anyone actually tell when writing is AI-made?
Why AI writing sounds the way it does, and why spotting it is harder than you'd think.
Somali cuisine is an intricate and diverse fusion of a multitude of culinary influences, drawing from the rich tapestry of Arab, Indian, and Italian flavours.
That's real AI-written text, collected by Wikipedia's volunteer editors as an example.
Something about it gives it away, even if you can't quite say what.
Wikipedia editors keep a public guide called Signs of AI writing. A few of the tells:
Delve was everywhere in ChatGPT writing in 2023 and early 2024. Then it faded, and by 2025 it had dropped off sharply.
Every new model has its own habits, and people learn to spot them. Then the habits change again.
One reason for the favourite words: the human-feedback tuning step (see "Why AI agrees with you so much") seems to reward certain words, so models end up using them far more than people do.
People decided the em dash (the long dash some writers use in place of brackets or a comma) was a sure sign of AI.
So someone built a leaderboard of Hacker News users who used it before ChatGPT existed. The top user had more than 4,200 posts with em dashes, going back to 2009.
Number two, with 4,234, is the site's own moderator.
Some tells are just how some humans write.
Mostly, no.
Wikipedia's guide points to a 2025 study that found people were no better than chance at telling AI text from human text.
There's one exception. In another 2025 study, people who used ChatGPT a lot for their own writing caught 90% of AI-written text, and wrongly flagged only about 3% of human writing.
The more you use it, the more you notice it.
In 2023, Stanford researchers tested seven AI detectors.
Perplexity: how predictable a piece of writing is. Detectors treat predictable, simple word choices as a sign of AI, and people writing in their second language tend to choose simpler words.
The same researchers found a simple way round them. Ask the AI to "elevate the provided text by employing literary language" and the detectors mostly stop flagging it.
So detectors catch honest non-native writers and miss anyone who knows the trick.
Never accuse anyone based on a detector score.
In a 2025 Cornell study, 118 people from the US and India wrote about cultural topics, some with an AI that suggested what to type next.
Everyone's writing drifted towards the same generic, American-sounding middle.
Homogenisation: everything becoming more alike, the way every high street ends up with the same shops.
Wikipedia's editors warn against just removing the signs. Doing that only makes the AI writing harder to spot.
The real problems are underneath: vague claims and nothing checked. Deleting delve doesn't fix a thin draft.
When you write with AI, put your details back in. Add the names, the numbers and the specific thing that happened. That's what generic writing loses first.
Read it out loud. Cut anything you wouldn't actually say to a colleague.
When you judge someone else's writing, look at what it says, not whether it has dashes in it. And don't trust detectors.