How to spot AI writing is less a matter of policing a dash than looking at patterns across a passage, according to The Economist’s July 30 analysis. Its central warning is useful: a conspicuous word, a long sentence or one piece of punctuation cannot establish who wrote a text. Model habits vary, and they change after software updates.
The magazine tested that proposition by asking OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini and xAI’s Grok to produce versions of its articles from AI-generated summaries, without web access. The Economist said it compared the resulting corpus with its own prose across 55,940 sentences and 1.2m words. It also checked the AI output against work from CNN, The New York Times and The Washington Post, plus excerpts from popular novels published between 1950 and 2022.
That is a study of tendencies in a particular collection of writing, not a detector that can settle an authorship dispute. Dr Karolina Rudnicka, a University of Gdańsk linguist who the university says contributed commentary to the piece, said there is no single AI style, just as there is no single human style.
What did The Economist find about how to spot AI writing?
The reported patterns cut against the most popular folk test. An em dash alone does not identify machine-generated prose. The Economist found that Claude was the only model in its comparison to use more em dashes than the human writers, while ChatGPT used markedly fewer than any writer studied. So banning a punctuation mark on sight is an odd way to do authorship analysis.
The magazine instead reported relatively sparse punctuation in the LLM text: fewer commas and semicolons, and scarcely any parentheses. It linked that result partly to longer sentences, with “and” identified as the models’ most overused word. The outputs also used more Latinate suffixes than the human material, a tendency The Economist connected to jargon-heavy or needlessly elevated wording.
Those features can help a reader decide where to look more closely. They cannot prove that a person used AI, or that a person did not. Context carries more weight: compare a suspect passage with the author’s established writing, and account for editing and the ways AI tools may have been used, as the University of Gdańsk’s account of Rudnicka’s comments advises.
The mechanism also explains why a durable signature is hard to pin down. Large language models generate text one token at a time, selecting likely continuations from learned patterns; changing the model or the software guiding it can change the visible prose habits. The practical conclusion is unglamorous but sound: treat stylistic tells as prompts for scrutiny, not forensic evidence.
This story draws on original reporting from Daring Fireball.
