The Experiment: Testing Human Writing for AI Tells
As artificial intelligence becomes increasingly integrated into content creation, writers face a new challenge: ensuring their work doesn't unintentionally mimic AI-generated text. A journalist recently put this concern to the test by submitting 10 recent articles—totaling roughly 11,700 words—to Claude Sonnet 4.6 for analysis. The goal was straightforward: identify any patterns that might make human writing sound robotic or AI-like.
The results were illuminating. While the overall assessment gave the writing a low AI-likeness score of 3 out of 10, several specific habits stood out as potential red flags. The most prominent was the overuse of parenthetical asides, followed closely by frequent em dashes and a reliance on filler words like "actually" and "rather." These patterns, though common in human writing, can trigger suspicion in both AI detectors and discerning readers.
Parenthetical Asides: The Number One Red Flag
Topping the list was the use of parenthetical asides—instances where additional information is tucked inside parentheses instead of being integrated into the main sentence. In the analyzed articles, this occurred 67 times across the sample, or roughly once every 175 words. This structural habit mirrors a common AI trait: hedging by adding caveats and clarifications without committing to a direct statement. Human editors often prune such asides, but writers who produce content quickly may leave them intact.
The reason parentheticals are so closely associated with AI is that language models are trained to maximize informativeness and minimize risk. By adding parenthetical notes, they can include qualifying details without disrupting the primary clause. Human writers, especially those accustomed to drafting quickly, may adopt the same approach—but it can make their prose feel overly cautious or mechanical.
Em Dashes and Other Structural Cues
Em dashes ranked second, with 78 instances across the sample—roughly one every 150 words. While em dashes are a legitimate stylistic tool for creating dramatic pauses or emphasizing key points, their density in human writing can mimic AI's tendency to splice clauses together rather than using periods or commas. In the analyzed writing, many dashes were used appropriately, but the overall frequency was high enough to raise eyebrows.
Other structural indicators included long sentences: 21% of sentences exceeded 35 words, with an average of 25 words. AI models often produce lengthy, technically correct sentences that cover multiple ideas at once, which can feel exhaustive rather than engaging. Human writers aiming for depth may inadvertently fall into the same pattern.
Filler Words and Hedges
The analysis also flagged specific words that appear disproportionately in AI-generated text. "Actually" appeared 15 times and "rather" 17 times—both acting as filler hedges. Similarly, "may" and "might" occurred 25 times combined, signaling a reluctance to make definitive claims. The impersonal "it is/was/has/seems" construction appeared 14 times, often at the start of sentences. Finally, "very" appeared 15 times as a generic intensifier that adds little meaning.
These lexical patterns are not inherently problematic, but when concentrated, they contribute to a voice that sounds tentative or formulaic. AI models favor hedging because it reduces the risk of factual errors, but human writers who adopt similar language may inadvertently echo that cautious tone.
What Makes Writing Sound Human
Despite these red flags, the overall assessment was reassuring. The writing scored low on the AI meter largely due to several humanizing factors. The vocabulary analysis showed no use of the overly complex or formal words that AI detectors often flag. First-person voice was strong, providing personal perspective and opinion. Sentence starters were varied and casual—using words like "but," "so," and "don't"—which deviates from AI's tendency to begin with formal transitions. Most importantly, the writer committed to opinions rather than hedging everything into neutrality.
The experiment underscores a key insight: while certain structural and lexical habits can mimic AI, a strong personal voice and natural sentence variety go a long way in preserving human authenticity. Writers who edit quickly and don't always tighten their prose may exhibit these tells, but that very inconsistency is a sign of human authorship. As AI detection tools become more sophisticated, the boundaries between human and machine writing will continue to blur, but the best defense remains a distinctive, opinionated style.
Source: PCWorld News