As AI-generated content becomes more common online, people are growing more vigilant about telling bot-written copy from genuine human writing. LinkedIn has even become the first major platform to add a button that lets users report "AI slop."
But how can readers be sure that what they are reading is actually the product of a large language model?
While there are some widely regarded tells of AI writing, those stereotypes may not be as reliable as they seem. A 2026 analysis by The Economist examined the state of AI writing, identifying patterns readers could look for as well as red herrings that do not actually point toward AI.
The analysis compared The Economist's published articles with versions of the same pieces generated by four leading AI models: OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and xAI's Grok. The sample also included writing from other news outlets, such as The New York Times and The Washington Post, as well as excerpts from popular novels published between 1950 and 2022. Altogether, The Economist analyzed 55,940 sentences and 1.2 million words.
What not to look for
Perhaps the most infamous sign of AI writing is its overuse of em dashes. Though they were long a favorite convention of writers, the rise of AI led many people to avoid the punctuation mark for fear that readers would assume their work was not human-made.
But according to the report, em dashes are no longer a reliable sign of AI-generated content. Of the four major models tested, only Claude used them more often than human writers.
In fact, a lack of punctuation may be a better indicator of AI-generated text, the report says. It found that the models used fewer commas, semicolons, and parentheses than humans. They also tended to craft overly long sentences, with "and" as their most overused word.
Some stereotypes ring true
AI-generated text is known for being unnecessarily wordy. The report found that large language models used rarer words and scientific language more often than humans. They also favored longer words and nominalizations, which turn verbs or adjectives into nouns, as when "nominalize" becomes "nominalization."
The models also relied on familiar rhetorical habits, such as the "it's not X, it's Y" contrast, which rejects one description before replacing it with another, and the rule of three. These habits shaped entire sentences and paragraphs: AI-generated writing tended to feature long sentences with little variation in length, producing blocky paragraphs with a uniform rhythm.
Another accurate perception of AI is that it learns fast, so the clues do not stay fixed. The models are constantly evolving and being trained on human writing, making the differences harder to spot. Not long ago, ChatGPT used em dashes heavily. Now it uses them less than any other model in the comparison and far less than human writers.
The rules for identifying AI-generated content are therefore always shifting. For now, the report suggests that verbose, punctuation-light writing and unusually uniform sentence structures may be more revealing than any single punctuation mark.