For years, journalists and content creators have faced a peculiar modern accusation: their writing sounds like artificial intelligence. The rise of large language models has made readers hyperaware of certain stylistic patterns, and many writers now wonder whether their natural voice has been unknowingly shaped by the very tools they cover. One tech writer decided to put this question to the test by submitting ten of his most recent articles to Claude, a leading AI assistant, asking it to identify every possible 'AI tell' in his prose.
The experiment aimed to move beyond vague online detectors and produce a concrete, ranked list of habits that trigger suspicion. The writer fed approximately 11,700 words to Claude Sonnet 4.6 using the Claude Cowork tool, with a simple prompt: list all the ways his writing sounded like AI, including frequency counts for classic tells. The results were illuminating, and the top offender was not what he expected.
The Top AI Tell: Parenthetical Asides
Claude found 67 parenthetical asides across the sample, making them the most frequent and structural red flag. Parentheticals are phrases or whole sentences inserted within parentheses, often used to add clarifications, exceptions, or asides. The AI noted that while long sentences and em dashes can be stylistic choices, the constant use of parentheses reads like hedging—a way of saying something while quietly footnoting the contradictions. Human writers who rely on this device so heavily are usually edited out of it by professional editors.
The writer acknowledged the habit, admitting that editors had called him out for it before. Parentheticals offer a quick way to sneak in extra information without committing to a full sentence, but they can make prose feel tentative and overly cautious. In the age of AI, where machines often produce hedging language, this pattern stands out as especially machine-like.
Em Dash Overuse: A Close Second
Coming in second were em dashes, with 78 instances—roughly one every 150 words. AI models frequently use em dashes as a crutch to splice clauses that should either be separate sentences or connected by a comma. While many of the writer's dashes were appropriate, the density was abnormally high. Em dashes add dramatic pauses and emphasis, but overuse can make writing feel choppy or unsure of its own structure.
The writer confessed a personal fondness for em dashes, calling them a favorite tool for emphasizing big points. However, the analysis confirmed that this affection, when shared by AI, becomes a telltale sign. Humans who lean too heavily on em dashes risk being mistaken for a language model trying to simulate natural flow.
Other Common AI Tells Revealed
Claude's full list included several other frequent patterns. Long sentences accounted for 21 percent of all sentences exceeding 35 words, with an average length of 25 words. AI tends to produce technically correct but sprawling sentences that try to cover too much at once, and the writer's sample showed a similar tendency.
The word 'actually' appeared 15 times, and 'rather' appeared 17 times—both filler hedges that AI (and AI-influenced writing) leans on heavily. At the writer's word count, 'actually' showed up once every 775 words. Additionally, the hedging words 'may' (14 occurrences) and 'might' (11 occurrences) added up to 25 instances. While not alarming individually, these choices contribute to a cautious tone that mirrors AI's default hedging behavior.
The classic impersonal opener 'It is' or 'It was' constructions occurred 14 times. Phrases like 'It's worth noting,' 'It's a snap to,' and 'It's clear that' are favorites of AI because they allow the model to make a claim without directly committing to a subject. Finally, the generic intensifier 'very' appeared 15 times, a word that adds little meaning and is easy to cut.
Despite all these patterns, Claude rated the writer's overall AI-likeness at only 3 out of 10. The AI explained that the vocabulary analysis was the key factor: the sample was completely clean on words that AI detectors and human readers actually flag. The first-person voice was strong, sentence starters were varied and casual (using words like 'but,' 'so,' and 'don't'), and the writer committed to opinions rather than hedging everything into mush. Parentheticals and em dashes would read as stylistic quirks to most people, not AI tells.
What This Means for Modern Writers
The experiment offers valuable lessons for anyone concerned about sounding like a machine. First, awareness of specific tells—overused parentheses, em dashes, filler words, and impersonal constructions—can help writers edit themselves more effectively. Second, the overall assessment shows that a strong personal voice and varied sentence structure outweigh many minor quirks. Writers who maintain a direct, opinionated style are far less likely to be mistaken for AI, even if they occasionally fall back on common patterns.
The journalist behind the test vowed to reduce his reliance on em dashes and parentheticals, acknowledging that these habits, while comfortable, may need to be curbed in an era where readers are suspicious of any hint of machine generation. He also noted that the exercise echoed a recent trend in writing advice: instead of asking AI to rewrite content, ask it to provide an honest critique. This approach preserves the writer's unique voice while benefiting from an external perspective on weaknesses.
Ultimately, the test confirms that human writing retains distinct markers that AI has not yet fully replicated. Natural variation in sentence length, colloquial expressions, and genuine opinionated phrasing remain hard for machines to imitate convincingly. Writers who fear sounding like AI should focus not on eliminating every possible tell, but on strengthening the elements that make their prose unmistakably human.
Source: PCWorld News