You can spot AI writing in five seconds. The em dashes, the tricolons, the throat clearing intros, the "underscoring the importance of." It is not subtle. It is so unsubtle that readers now bounce the moment they see it, and the trust you spent years building gets undercut by output you posted in two minutes.
The cause is not that AI cannot write well. The cause is that AI defaults to a specific style, and most people never override the defaults. The defaults are obvious. The overrides are not.
This post breaks down seven specific tells that mark text as AI generated, why each one happens, and the prompt pattern that removes it. Apply the fixes and your output stops broadcasting its origin. Apply them consistently and your readers stop being able to tell.
Tell 1: Em Dash Overuse
The em dash is the single most reliable AI tell. Models love it because em dashes signal sophistication without committing to a specific punctuation rule. Most human writers use em dashes occasionally. AI uses them in almost every paragraph, often more than once.
What it looks like. "The launch was successful, however, we learned a few lessons along the way. Some of them, including the most important one, came from unexpected places, like the support inbox, which we had previously underweighted."
The fix. Add a negative specification to your prompt. "Do not use em dashes. Do not use hyphens to join phrases. Use periods, commas, or parentheses instead."
The same content, rewritten under the constraint, reads as written by a human who chose simpler punctuation. The voice is more direct, less performatively literary.
Tell 2: Rule of Three Overuse
AI loves triplets. Every sentence becomes "X, Y, and Z" or "fast, scalable, and reliable" or "we delivered insights, alignment, and outcomes." Once you see the pattern you cannot unsee it. Real human writing varies between pairs, singletons, and longer enumerations. AI writing rarely does.
What it looks like. "Our approach is fast, scalable, and reliable. We help teams plan, execute, and measure. The platform supports collaboration, automation, and reporting."
The fix. Add an explicit constraint on the pattern. "Vary the rhythm of your enumerations. Use pairs more often than triplets. Use singletons where one word would do. Do not write three sentences in a row that each end in a list of three."
The output starts to sound like someone actually thinking, not a model reaching for a familiar cadence.
Tell 3: Inflated Significance Language
AI writes about everything as if it matters more than it does. Phrases like "underscores the importance of," "represents a pivotal moment," "highlights the critical role," and "serves as a testament to" show up everywhere in AI text because the training data rewards them. Real writers rarely use any of these, especially in casual or operational contexts.
What it looks like. "The new pricing model represents a pivotal moment for our team, underscoring the critical importance of clear communication and serving as a reminder of how aligned execution drives outcomes."
The fix. Use negative specification with a target list. "Do not use these phrases or any variation: underscores the importance of, represents a pivotal moment, highlights the critical role, serves as a testament to, marks a turning point, plays a key role, reflects a broader trend. State the fact directly without inflation."
The rewrite becomes almost shorter than the original. "The new pricing model is live. We learned that clear communication matters more than we expected."
Tell 4: Negative Parallelism
The "it is not just X, it is Y" construction is one of the most overused rhetorical moves in AI text. Models reach for it constantly because it sounds insightful. In practice it usually adds nothing and makes the writing feel performative.
What it looks like. "This is not just a product launch. It is a statement about who we are. This is not just about features. It is about how those features change the workflow. This is not just a tool. It is a partner in your daily work."
The fix. "Do not use the not just X but Y construction. Do not use chiastic reversals. Make the point once and move on."
The rewrite gets directly to the point. "The launch matters because it changes how the workflow runs day to day."
Tell 5: Vague Universal Openings
Every AI generated article seems to start with "In today's fast paced world" or "In an era of unprecedented change" or "As businesses increasingly look to" or "Now more than ever." These openings are content free. They exist because the model is buying time before it commits to a specific claim.
What it looks like. "In today's rapidly evolving business landscape, organizations are increasingly looking to artificial intelligence to drive efficiency, unlock new capabilities, and stay ahead of the competition."
The fix. Add an opening constraint. "Start the piece with a specific claim, observation, or example. Do not start with phrases like in today's, in an era of, now more than ever, as the world becomes, or any variation. The first sentence must be content not throat clearing."
The rewrite forces a real opening. "Most companies that adopted AI in 2024 reduced report generation time by half. Most also produced text nobody wanted to read."
Tell 6: Adjective Stacking
AI marketing writing piles adjectives. "Seamless, intuitive, and powerful." "Robust, scalable, and innovative." "Cutting edge, comprehensive, and user friendly." Each adjective adds nothing because each one is vague. Real writing chooses one specific descriptor or skips the adjective entirely.
What it looks like. "Our seamless, intuitive, and powerful platform delivers cutting edge insights through a robust, scalable, and user friendly interface."
The fix. Use a banned word list. "Do not use these adjectives: seamless, intuitive, robust, scalable, innovative, cutting edge, comprehensive, user friendly, world class, best in class, state of the art, next generation. Replace each adjective with either a specific verifiable claim or remove it."
The rewrite becomes lean. "Our platform shows you which prompts produced the highest open rates in the last 30 days." A real claim is worth more than five vague adjectives.
Tell 7: Hollow Positive Closings
AI almost always ends on an upbeat note. "Exciting times ahead." "The future looks bright." "We are just getting started." "The journey continues." These closings exist because the training data ends in upbeat conclusions, and the model reaches for them by default. Human writing often ends mid thought, with a question, or with a small specific observation, not with a slogan.
What it looks like. "As we look ahead, the possibilities are limitless. We are just getting started, and exciting times lie ahead. The future is bright, and we cannot wait to see what comes next."
The fix. Constrain the ending explicitly. "Do not end with a generic positive conclusion. Do not use phrases like exciting times ahead, the future looks bright, we are just getting started, the journey continues, or anything similar. End with a specific observation, an open question, or just stop."
The rewrite has a real ending. "Three of our biggest wins last quarter came from prompts written under 100 words. The rest came from prompts we rewrote three times."
How to Stack the Fixes Into One Prompt
Each fix above works on its own. Stacked together they remove the AI fingerprint almost completely.
A working anti AI tell preamble looks like this. "Write in a direct, human voice. Do not use em dashes. Do not use the not just X but Y construction. Do not use these phrases: underscores the importance of, represents a pivotal moment, in today's fast paced world, exciting times ahead, the future is bright. Do not stack adjectives, especially seamless, robust, intuitive, scalable, innovative, cutting edge. Vary sentence length. Use pairs and singletons more often than triplets. Start with a specific claim, not a setup phrase. End with a specific observation, not a slogan."
Paste that as the first paragraph of any writing prompt. The output changes immediately. It still has minor tells, but the obvious ones are gone, and the text reads as drafted by a person who happens to be using AI as a tool, not as AI pretending to be a person.
Why This Matters For Anyone Publishing AI Assisted Content
The fingerprints above are not just stylistic preferences. They are now active signals readers use to decide whether to trust the writing. Once a reader spots three or four of the tells in a single paragraph, they pattern match and bounce. The piece could be excellent on substance and still lose the reader on style alone.
The economic argument is simple. Most writers using AI today take the default output. The minority who add a humanization preamble produce content that reads as theirs. As detection improves on the reader side, the default output gets penalized in search, social, and direct reader trust. Removing the seven tells is no longer a stylistic choice. It is the minimum bar for usable AI assisted writing.
Frequently Asked Questions
Why does AI default to these specific patterns?
Large language models learn from text that overweights formal, persuasive, and promotional registers. The seven tells above are common in marketing copy, op eds, and corporate communications, all of which appear heavily in training data. The model is not making stylistic errors. It is doing exactly what its training distribution rewards.
Do these fixes work for ChatGPT, Claude, and Gemini?
Yes. The fixes target the model's output behavior, not any specific provider's quirks. The same anti tell preamble works in any modern instruction following model.
Will using these fixes make my AI writing perfect?
No. It removes the most obvious tells. Subtler patterns remain, including paragraph rhythm, transition consistency, and tonal evenness. Removing the seven above gets you 80 percent of the way to natural sounding output. The remaining 20 percent comes from light human editing.
Should I edit AI output instead of using better prompts?
Both. A better prompt reduces the editing time. Editing catches what the prompt missed. Most professional AI assisted writers use both. The prompt does the heavy structural work. The edit catches the last few tells and adds specific facts the model could not have known.
Are em dashes always wrong?
No. Em dashes are fine in moderation. Real human writers use them. The problem is volume. AI uses em dashes at roughly five times the rate of human writers. Removing them entirely is the easiest fix. Using them rarely is the more sophisticated one.
Is there a tool that detects these tells automatically?
A few exist, with varying accuracy. The fastest detector is a human reader who has seen enough AI output. The seven tells in this post are the ones that consistently trigger that pattern recognition. Removing them passes most human and tool based checks.
Can I use these patterns to humanize text someone else wrote with AI?
Yes. The same fixes work as editing rules. Read through the text, find each tell, rewrite it. The seven categories above cover the majority of cases. A 500 word piece can usually be humanized in five to eight minutes of editing.
Get the Full Humanization Prompt Pack
The seven fixes above are the structural starting point. The PromptLeadz Humanization Prompt Pack expands them into 30 plus working prompts that rewrite, audit, and humanize any AI assisted writing in seconds, every prompt formatted three ways for Claude, ChatGPT, and Gemini. Built for writers, marketers, founders, and anyone shipping AI assisted content who cannot afford to sound like a model.
Browse the PromptLeadz role packs and Freebie Vault for more frameworks, prompts, and free starter packs in every category.
Zostaw komentarz: