ChatGPT and Claude reach for em dashes 6 to 8 times more often than human writers. Once a reader notices the pattern, they assume the rest of your content is machine-made too. The em dash is the loudest AI tell in 2026, but it sits on top of a deeper stack of word choices, cadences, and paragraph shapes that give AI writing away. Here is the full list, why it costs you buyer trust, and how to strip it out.
Why every AI model loves the em dash
Three years into the LLM era, one punctuation mark has become the unofficial logo of AI writing. It is the em dash.
The cause is simple. Large language models predict the next likely token based on what they read during training. Books, long-form essays, and serious nonfiction use em dashes often. Casual web writing does not. The models inhaled a punctuation pattern that reads as confident, literary, and edited, then applied it to every output: marketing emails, FAQ pages, LinkedIn posts, product descriptions.
The result is a stylistic fingerprint. A 2025 analysis by Plagiarism Today and a follow-up by MDPI both found that LLM outputs use em dashes at roughly six to eight times the rate of nonprofessional human writing in the same genre. ChatGPT in particular drops them where humans would use a comma, a colon, parentheses, or a simple period.
That is the actual giveaway. Not the em dash itself. The em dash placed where no human would have used one.
The vocabulary tells: words that scream AI
The em dash is the easiest pattern to spot, but it is far from the only one. Editors and readers who have been swimming in AI content for two years now recognize a deeper set of signals.
The GPT-4 era seeded a generation of words that have become near-synonyms for “this was written by a robot”: delve, tapestry, meticulous, pivotal, robust, navigate, realm, landscape, foster, showcase. The GPT-4o era added its own batch: leverage, utilize, holistic, dynamic, seamless, comprehensive.
None of these words are wrong in isolation. The problem is frequency. A 2026 analysis of 50,000 web pages by Bloomberry found that “delve” appears 47 times more often in AI-generated content than in human-written content from the same period. “Tapestry” appears 89 times more often. Each word functions as a fluency placeholder, signaling effort while skipping the harder work of saying something specific.
The hedging tells are just as loud. AI cannot risk being wrong, so it softens everything. Watch for phrases like “it is worth noting,” “it is important to consider,” “various factors can influence,” “many experts agree.” Human writers either commit to a claim or skip it. AI splits the difference.
The cadence tells: rhythms that only AI uses
AI loves a three-beat rhythm. Watch for sentences in this shape:
“It is not just about X. It is about Y. It is about Z.”
“This is more than a strategy. It is a mindset. It is a philosophy.”
“We do not just write content. We craft narratives. We build connections.”
Real humans write in three-beat patterns occasionally. AI does it constantly, because the pattern feels balanced to a token predictor without requiring any actual argument.
The same goes for the “more than” construction. “More than just a law firm.” “More than just a clinic.” Any time a piece of marketing copy frames itself in opposition to a lesser version of itself, a model probably wrote it.
The structural tells: how AI builds paragraphs and lists
Beyond word choice, AI gives itself away through structure.
Paragraphs run at identical length. AI tends to produce paragraphs that all land between two and four sentences. Human writers vary aggressively: a one-sentence paragraph for emphasis, then a six-sentence paragraph, then a fragment.
Lists come back with parallel construction so clean it looks edited. AI bullet points usually start with the same part of speech, run the same length, and end with the same kind of punctuation. Human bullet points are messier because humans get bored halfway through and shift form.
The mandatory conclusion is another tell. AI almost never ends a piece without a wrap-up paragraph that begins “In conclusion,” “Ultimately,” or “At the end of the day.” Real writers often just stop.
The sandwich opener completes the set. AI introductions follow a fixed template: establish the topic in one sentence, list two or three reasons it matters, then promise what the rest of the piece will cover. Once you see the pattern, you cannot unsee it.
The hidden cost: what AI-flavored content does to buyer trust
For a SaaS company or a content mill, AI tells are a brand annoyance. For a law firm, a cosmetic surgeon, or a financial advisor, they are something worse. They are a trust killer.
Buyers in high-stakes service categories read content as a signal of professional judgment. A managing partner researching firms for a complex commercial dispute does not just want information. They want to see how a firm thinks. A patient researching rhinoplasty surgeons wants evidence the surgeon has done it 800 times and can explain why this case is different from the last 100.
The moment a buyer reads “we leverage a holistic approach to navigate the complex landscape of estate planning,” the firm has signaled the opposite of what it intended. Not “we are sophisticated.” But “we did not write this, and we may not have thought about it either.”
This is the AEO problem too. When a buyer asks ChatGPT “what should I look for in a cosmetic surgeon for breast reduction,” the engine pulls from your site. If your site reads like a different LLM wrote it, the citation does not help. It creates an echo chamber of generic AI text recommending generic AI content, and the engines learn to weight it less.
What Google’s February 2026 core update revealed
Google has been clear since 2023 that it does not penalize content based on how it was produced. The February 2026 core update reinforced that policy and also showed where the actual penalty line sits.
Sites that published AI content at scale without editorial review lost 40 to 60 percent of organic traffic during the update. Semrush Sensor readings hit 9.4 during the rollout, one of the highest volatility scores on record. The pattern was consistent. Thin AI content with strong AI tells got demoted. AI-assisted content with human editing, original data, and specific opinion held rankings or gained.
Google is not detecting AI through em dashes. It is detecting it through behavior: do users bounce, do pages get cited, does the content add information that does not already exist on the web. AI tells happen to correlate with all three failure modes, because writing that reads like ChatGPT usually has no original information in it.
The SubscribePR editorial kill list
We publish content for law firms, cosmetic clinics, financial advisors, and other high-trust service businesses. Every article ships through a filter that strips AI tells out before publication. Here is the working list of patterns we never let through.
Punctuation: no em dashes. A comma, a period, a colon, or parentheses every time.
Banned words: robust, comprehensive, innovative, leverage, utilize, cutting-edge, game-changing, holistic, seamless, dynamic, navigate (as a verb), foster, showcase, delve, tapestry, realm, landscape, pivotal, meticulous, align with.
Banned phrases: “it is worth noting,” “in today’s digital landscape,” “in the realm of,” “at the end of the day,” “in conclusion,” “ultimately,” “needless to say,” “let us unpack,” “move the needle.”
Banned cadences: the three-beat “not just X, but Y, but Z” pattern. Sandwich introductions that telegraph the structure of the piece. Conclusions that summarize what the reader just read.
Required edits: every paragraph either makes a specific claim, cites a specific number, names a specific publication or tool, or shares a specific opinion. If a paragraph would survive a find-and-replace of “law firm” with “dental clinic,” it does not say anything.
How to rewrite AI drafts so they read human
We still use AI in the drafting process, both ChatGPT and Claude. The work is in what happens after the draft.
Read the draft and circle every em dash, every banned word, and every sentence over 25 words. AI cannot help itself with long sentences either.
Rewrite each paragraph in two sentences if the AI used three, and three if the AI used two. Vary length on purpose.
Add at least one piece of information per section that the AI could not have generated. A real case. A number from your own work. A publication name the AI would not know to cite. A specific opinion the AI would not have risked.
Cut the introduction in half. AI intros are 30 percent longer than they need to be.
Kill the conclusion. End on the last useful sentence.
A 1,500-word AI draft typically becomes a 1,100-word human article after this process. The reading time gets shorter, the citation rate gets higher, and the trust signal gets stronger. Buyers can feel the difference, even when they cannot articulate it.
Frequently asked questions
Are em dashes a reliable AI tell on their own?
No. Em dashes have legitimate uses in human writing, and any AI tool can be told to avoid them. The signal is not the em dash itself, but the frequency and the placement. AI uses them six to eight times more often than humans, and drops them in spots where a comma or period would do the same work.
Will AI detection tools catch ChatGPT or Claude output reliably?
Sometimes. GPTZero and Originality.ai both claim 96 to 99 percent accuracy on pure AI text, but independent benchmarks show 15 to 23 point gaps between vendor claims and real-world performance. Detection drops sharply on newer models. Originality.ai catches only about 32 percent of GPT-5 output. False positives also matter. GPTZero flags about 1 in 400 human documents as AI. Originality.ai flags about 1 in 20.
Does Google penalize AI-written content directly?
No. Google’s official position, restated through the February 2026 core update, is that it does not care how content is produced. It cares about quality, originality, and user intent. The sites that lost 40 to 60 percent of traffic during that update were running scaled, low-effort AI content with no editorial review. Sites using AI as a drafting tool with human editing held their rankings.
Can I just find-and-replace em dashes to fix the AI tells?
That fixes the easiest signal. It leaves the vocabulary, cadence, and structural tells intact. A reader who has seen a few hundred AI articles can still spot one from a paragraph away, even with the em dashes stripped out.
Should my law firm or clinic disclose that we use AI for content?
You do not have to, and most firms do not. The relevant standard is editorial. If a human professional reviewed the content, can stand behind the claims, and added something the AI could not have generated, that is the trust standard your buyers care about, not the tooling.
Will buyers in high-trust service categories actually notice?
Yes. Law firm clients, cosmetic surgery patients, and financial advisory prospects spend more time on your site before contacting you than buyers in lower-stakes categories. They are reading for tone and judgment, not just information. AI tells register as carelessness, which registers as risk.
The takeaway
If your firm publishes content that sounds like every other firm’s content, the AI engines will treat it that way. SubscribePR builds editorial systems that strip AI tells out before publication, keep your site cite-worthy in ChatGPT and Google AI Mode, and protect the trust signal that high-value buyers actually care about. See what your AI search visibility looks like today at our ROI calculator, or book a call to talk through what your firm’s content sounds like right now.
Tagged