What could be better than an em dash? They pepper my writing, adding dramatic pauses and emphasizing my biggest points. I love em dashes — so does AI.

The use — and overuse — of em dashes is, of course, a classic AI writing tell, along with such AI writing quirks as the “it’s not just this; it’s that” construction (which itself can be combined with an em dash) and lengthy sentences that seem to stretch into infinity. Like plenty of other online writers, my articles are regularly accused of sounding like AI. While web-based AI detectors have repeatedly assured me that my stories sounded human, I wanted more details — and most of all, I wanted to discover if my em-dash obsession was the writing crutch that drew the most “you’re AI!” fire.

For the test, I turned to Claude, feeding 10 of my most recent articles to Claude Sonnet 4.6 (using the Claude Cowork tool) and giving it a simple prompt: “Can you look at my writing samples in the directory and list all the ways in which I sound like AI, including how often I use classic ‘AI tells’ in my writing? Give me a ranked list.” I braced myself for the bad news, anticipating that overused em dashes (as I previously mentioned) would be my biggest — if false! — AI tell. Turns out, not quite. Here’s what Claude ranked as my number-one “sound like AI” issue: the parenthetical asides — 67 in ~11,700 words. It’s the one pattern that’s both frequent and structural. Long sentences can be a rushing problem, em dashes can be a style choice, but constantly tucking clarifications into parentheses instead of committing to the sentence reads like AI hedging its bets — saying the thing while also quietly footnoting the exceptions. Human writers who do it this much usually get edited out of it.

Yep, it’s true, I’m a sucker for a good parenthetical, and it’s a habit my editors have (quite rightfully) called me out for. (I know, I’m doing it again.) But rest assured, em dashes came in a close second: Em/en dash overuse: 78 instances — roughly one every 150 words. AI uses em dashes as a crutch to splice clauses that should either be separate sentences or use a comma. Many of yours are fine, but the density is high.

And here’s the rest of the list:

• Long sentences: 21% of your sentences exceed 35 words, and your average is 25 words. AI tends to write long, technically-correct sentences that feel like they’re covering too much at once.
• “Actually” x15, “rather” x17 — both are filler hedges that AI (and AI-influenced writing) leans on. At your word count, “actually” shows up once every 775 words.
• “May” x14 + “might” x11 = 25 hedges — not alarming, but AI tends to hedge rather than commit. These add up.
• “It is/was/has/seems…” constructions: 14 times — the classic impersonal opener. Phrases like “It’s worth noting,” “It’s a snap to,” “It’s clear that” — AI loves starting sentences this way instead of making a direct claim.
• “Very” x15 — generic intensifier that doesn’t add meaning. Easy to cut.

Guilty as charged. The good news, according to Claude, is that my writing generally does sound human, ranking just 3 out of 10 on the “sounds like AI” meter. Claude elaborated on my score, tossing in a few em dashes of its own for good measure: The vocabulary analysis is the main thing — you’re completely clean on the words that AI detectors and human readers actually flag. First-person voice is strong, sentence starters are varied and casual (“but,” “so,” “don’t”), and you commit to opinions rather than hedging everything into mush. The parentheticals and em dashes would read as stylistic quirks to most people, not AI tells. A detector might ding you for sentence length, but a human reader wouldn’t think “AI wrote this.” If anything, your writing reads more like a journalist who edits fast and doesn’t always tighten — which is a very human problem.

While I’m pleased to hear I mostly don’t sound like AI, Claude’s last point is a fair one – indeed, it echoes last week’s “critique, don’t rewrite” prompt, which makes an AI give you an honest evaluation of a writing sample. In any event, I’ll try using those em dashes more sparingly in the future, and the same goes double for parentheticals. That’s not just wishful thinking; it’s a promise.

To understand why these patterns matter, it helps to look at how AI language models generate text. Models like GPT-4 and Claude are trained on vast corpora of human writing, but they tend to over-represent certain stylistic features that appear more frequently in formal, edited prose. For example, em dashes are used in journalism to create emphasis, but AI models use them far more pervasively because they can easily connect ideas without needing full sentence breaks. Similarly, parentheticals allow a model to insert caveats and clarifications without breaking the flow, a behavior that mimics the hedging language common in academic and technical writing. The result is a style that feels both polished and cautious — sometimes too cautious.

The rise of AI writing detectors has made many writers self-conscious about their natural habits. Services like Originality.ai, GPTZero, and Writer.com offer scores that claim to distinguish human from machine. But these tools are often unreliable, especially for experienced journalists who have rigorous editing standards. In my case, running the same articles through three different detectors gave me readings ranging from 1% AI probability to 45%. The inconsistency highlights a fundamental truth: AI detection is an emerging field, and no tool is perfect. The better approach is to cultivate a voice that reads as authentically human — one that includes personal anecdotes, opinion, irony, and occasional grammatical looseness.

My ranking from Claude is a useful self-diagnosis. The parenthetical aside problem is something I’ve known about but never quantified. Seeing 67 instances in 11,700 words is sobering. It suggests that roughly one in every 175 words is a parenthetical aside. That’s high by any standard, whether human or AI. Editors typically recommend using parentheticals sparingly — no more than one per paragraph, and only when the aside is genuinely additive. Otherwise, the writing feels fragmented and indecisive. I plan to review my recent drafts and cut at least half of those asides, either integrating the information into the main clause or deleting it altogether.

The em dash overuse is similarly telling. With 78 instances in the same word count, that’s one dash every 150 words. In a typical 800-word article, that would be about five em dashes. Some are fine, but five is excessive. The rule of thumb is to limit em dashes to no more than one per paragraph, and only for dramatic effect or abrupt interruption. Replacing many of them with commas, semicolons, or separate sentences would make the writing stronger and less AI-like.

Long sentences are another issue. AI models often produce long, grammatically perfect sentences that can run on for 50 words or more. My average sentence length of 25 words is within the normal range for journalistic writing (which typically averages 20-30 words), but the fact that 21% exceed 35 words is a red flag. The solution is to vary sentence length deliberately, using short, punchy sentences for emphasis and longer sentences for explanation. This creates a rhythm that feels more natural and less mechanical.

The filler words “actually” and “rather” are both subtle signals of uncertainty. “Actually” often implies a correction or surprise, but when overused, it makes the writer seem as if they’re constantly contradicting expectations. “Rather” is a formal hedge that can be replaced with “instead” or “more accurately.” Reducing these from 15 and 17 occurrences to just a handful each would tighten the prose significantly.

Hedging with “may” and “might” is a hallmark of cautious writing. While some hedging is necessary in technology journalism where speculation is common, 25 instances across 11,700 words is borderline. AI models tend to over-hedge because they are trained to avoid absolute statements. Human writers can take a stand more often, using phrases like “will” or “does” when the evidence is strong. Similarly, impersonal openers like “It is worth noting” or “It seems that” create distance. Starting sentences with a strong subject or a concrete observation is more engaging.

Finally, the generic intensifier “very” appears 15 times. This is a classic redundancy that adds little. Replacing “very” with a stronger adjective (e.g., “very important” → “crucial”) or simply cutting it can improve clarity and reduce wordiness.

All these patterns, while individually minor, add up to a style that can trigger AI detectors and, more importantly, bore human readers. The best defense against sounding like AI is to read your work aloud, ask for human feedback, and deliberately vary your sentence structure and word choices. The goal is not to eliminate every em dash or parenthetical, but to use them with intention rather than habit.

In my case, Claude’s analysis was a wake-up call. I now have a concrete list of habits to address. The process of self-editing will be ongoing, but the insights from this test are invaluable. I encourage every writer, whether or not they worry about AI detection, to run a similar analysis on their own work. It’s a way to see your blind spots and grow as a craftsman. And if you do, try to avoid the em dashes — at least until you’ve counted them.

Source:PCWorld News