How One Reviewer Spots AI-Written Papers Without Any Tools

One academic journal reviewer has now read through thirty AI-generated paper submissions this year. The thirty-first invitation arrived in their inbox today. The reviewer says no detection software is needed. You can just feel it. That sounds vague, but the specifics are clear. AI-generated academic papers share a recognizable set of patterns: adjectives are excessively glowing ("groundbreaking," "intricate"); made-up theoretical terms appear without foundation; each section opens by announcing what it is about to do, as if the author is checking notes rather than addressing a reader; identical sentences show up three or four times across the paper; and em-dashes appear far more often than natural writing calls for. Any one of these can appear in ordinary writing. When several show up together, the feeling is immediate. The same patterns appear in English. Researchers developing tools to strip away AI writing's telltale signs found that English AI-generated text reliably favors phrases like "not just X, but Y" and "it is worth noting that." AI writing in both Chinese and English uses similar devices to make the text look like careful thinking. Tools now exist to clean up these signals: remove the excessive adjectives, swap out the filler phrases, restructure the sentences, bring AI detection scores down. Once the tools can do something, the detectors quickly fall behind. But stripping the signals does not settle the deeper question. That reviewer notes that even after those surface patterns are removed, AI-generated papers in the social sciences still tend to contribute less. Social science research depends on interpretation, argumentation, and the mapping of cause and effect. Stacking words together does not get you there. A paper can be correct on every sentence and still feel like something is missing. That feeling might be the trace of someone who genuinely worked through a problem. Something got thought about, went sideways, and got rethought. That process leaves marks in writing. The tools, so far, are not very good at faking those marks. Research found that of 186,000 news articles published between June and September 2025, about 9 percent were partly or fully AI-generated. Academic papers are almost certainly higher. The arms race between generators and detectors is not close to over. The thirty-first paper arrived in the inbox today.