How to Humanize Your AI Writing

1. Generic AI output is a math problem, not a mystery. Models are trained to produce the most probable next word, which means they regress toward the statistical average of billions of pages of corporate blogs, news, and marketing copy. The result is prose that is competent, smooth, balanced — and anonymous. It contains everyone's voice, which means it contains nobody's.

2. There is now an authoritative, free taxonomy of the "tells." Wikipedia's WikiProject AI Cleanup published "Signs of AI writing," a roughly 15,000-word field guide built from thousands of real examples. It sorts the patterns into buckets: Language and tone, Style, Markup, Citations, and Communication intended for the user. Editors single out two patterns as the strongest tells: "undue emphasis on symbolism and importance" and "superficial analyses."

3. The fix is two-part: subtract the robot, add yourself. Removing tells (em dashes, jargon, hollow structure) only gets you to "neutral." To get to "sounds like me," you must feed the model samples of your own writing and define your voice. The best practitioners combine both: a "what NOT to sound like" file plus a "this is how I sound" file.

4. Don't over-ban. If you strip out every possible human-sounding device, the writing becomes stiff and more obviously artificial. Prompt out a few high-signal patterns (em dashes, negative parallelism), then teach voice with examples.

5. There's a real professional stakes argument. A 2025 University of Florida / USC study of 1,100 professionals found that only 40%–52% of employees viewed supervisors as sincere when they used high levels of AI, versus 83% for light-assistance messages. So "humanizing" is not vanity — it protects credibility.

If you have never used ChatGPT, Claude, or Gemini to write something, here is the one thing to understand before anything else: these tools do not "know" things and do not "think" about how to phrase an idea. They are prediction engines. Given the words so far, they calculate the most statistically likely next word, then the next, and so on.

"Most likely" means "most common in the training data." That training data is billions of words scraped from the internet — corporate blogs, press releases, news articles, Wikipedia, marketing pages, forum posts. When you type "write a blog post about leadership," the model produces something close to the average of every leadership blog post it has ever seen. Not terrible. Not brilliant. Just average. The same phrasing everyone uses, the same balanced structure, the same personality-free tone.

Wikipedia's editors describe this beautifully. AI "tends to regress to the mean," they write, sanding down "specific, unusual, nuanced facts" and replacing them with "generic, positive descriptions." A precise detail like "invented a train-coupling device" becomes "a revolutionary titan of industry." As the guide puts it, the result "is like shouting louder and louder that a portrait shows a uniquely important person, while the portrait itself is fading from a sharp photograph into a blurry, generic sketch. The subject becomes simultaneously less specific and more exaggerated."

That is the whole problem in one sentence. The model defaults to vague, flattering, polished, balanced filler — because that is the safest statistical bet. Your job as the human is to override that default.

There is also a real cost to getting this wrong. A 2025 University of Florida and University of Southern California study published in the International Journal of Business Communication (Cardon & Coman, DOI 10.1177/23294884251350599) surveyed 1,100 professionals about AI-assisted manager emails. The trust drop was steep: "Only 40% to 52% of employees viewed supervisors as sincere when they used high levels of AI, compared to 83% for low-assistance messages." Perceived professionalism fell too — 95% rated low-AI messages as professional, but that dropped to 69–73% for high-AI messages. Co-author Anthony Coman of UF's Warrington College of Business warned: "Despite positive impressions of professionalism in AI-assisted writing, managers who use AI for routine communication tasks put their trustworthiness at risk when using medium- to high-levels of AI assistance." Readers can feel the average-ness, and they hold it against you.

The most important source for any beginner is Wikipedia's "Signs of AI writing" guide, maintained by volunteer editors who clean up AI-generated articles. It is descriptive, not a rulebook. Their own disclaimer matters: "This list is not a ban on certain words, phrases, or punctuation. Not all text featuring these indicators is AI-generated, as the large language models that power AI chatbots are trained on human writing, including the writing of Wikipedia editors." They even add the joke: "No one is taking your em-dashes away." In other words, humans use these devices too. These are probabilities, not proof — "potential signs of a problem, not the problem itself." But if you want to avoid sounding like AI, this is your checklist. LinkedIn creator Ruben Hassid built his entire anti-AI method on copying this exact Wikipedia page into a file.

Here are the major categories and the specific tells.

Language and tone (the most important bucket):

Style and formatting:

Markup and citations: leftover Markdown asterisks, fill-in-the-blank placeholders like "[insert detail here]," hallucinated or broken citations, and fabricated sources (invalid DOIs and ISBNs).

Communication intended for the user: the dead giveaways someone forgot to delete — "Certainly!", "I hope this helps," "Would you like me to…", "As an AI language model…", "Here is your post:".

A useful corroborating data point: when Leadership IQ analyzed real workplace writing pasted into its AI detector, it concluded the biggest tell was not vocabulary at all but sentence-length variation. Per its report (via Yahoo, "Why AI Writing Is Quietly Eroding Your Professional Credibility"): "In writing flagged as human-generated, more than a quarter of sentences were five words or shorter. In AI-flagged writing, fewer than four percent were that short. Humans use fragments." Human writing also used exclamation points, question marks, and parenthetical asides far more often, and opened sentences with "I," "So," and "You" rather than "It," "This," and "In."

Removing tells gets you to neutral. To sound like you, you have to give the model raw material about how you actually write. Here is the workflow, simplest first.

Technique 1: Feed it samples (the single highest-leverage move). Gather three to five pieces of writing that genuinely sound like you — emails, a past article, a memo, even transcribed voice notes. Paste them in and say: "Here are samples of my writing. Analyze my tone, sentence length, vocabulary, pacing, and personality. Then write [your task] in that same voice." Almost every credible practitioner — Sabrina Ramonov, Zapier's guide, Forte Labs, Futurepedia — converges on this "priming" step. The model is extraordinary at mimicking patterns when you give it something to copy.

Technique 2: Build a reusable voice profile. Rather than re-pasting samples every time, ask the AI to turn its analysis into a saved "style guide" — a document describing your voice, your favorite words, words you'd never use, sentence shapes, and how you open and close. Save this. In ChatGPT, paste it into Settings → Customize ChatGPT ("What traits should ChatGPT have?"). In Claude, drop it into a Project's instructions. Now every chat starts in your voice. (Note: ChatGPT's custom-instructions box has a ~1,500-character limit, so you'll trim to essentials.)

Technique 3: The interview method (Ruben Hassid's "Taste Interviewer"). Instead of describing your own voice — which most people do badly — have the AI extract it. Hassid's prompt instructs the model to act as a "relentless interviewer" and ask 100 one-at-a-time questions across categories like beliefs and contrarian takes, writing mechanics, "aesthetic crimes" (what makes you cringe), voice and personality, structural preferences, hard nos, and red flags. It pushes back on vague answers ("Simple how? Give me an example of simple done right and simple done lazy"). At the end it produces a compact ".md" voice file you can reuse in any AI tool. Hassid's tip: dictate your answers by voice, because "voice is faster and more honest."

Technique 4: The two-file system. Hassid's core insight is to separate two jobs. One file says what not to sound like (the "anti-ai-writing-style" file — essentially the Wikipedia tells plus a banned-word list). A second file says who you are (your voice profile). He argues a short prompt plus a thorough file beats a giant 500-word prompt, because a wall of "don'ts" gets forgotten "by sentence three," whereas a file the model reads and audits against actually sticks: "The file does the work. Not the prompt." His everyday prompt is tiny — "I want to [task] for [success criteria]" — followed later by "Audit your text against the anti-ai-writing-style file."

Technique 5: The editing pass (never skip this). Even great prompts leave artifacts. Sabrina Ramonov is blunt: "Don't copy paste from ChatGPT without humanizing the text first. Otherwise, you will sound like AI slop and turn people off from understanding the ideas you were trying to communicate." Read the draft aloud. Cut the throat-clearing first sentence. Delete every hedge. Replace one vague claim with a specific number or name. Break up three same-length sentences. Add one genuine reaction or opinion the AI could never have. This takes two minutes and does most of the work.

Sabrina Ramonov (AI educator, founder of Blotato) treats the humanize prompt as a starting point, not a magic fix. Her approach: append a writing-style block to your prompt, then save a trimmed version to ChatGPT's custom instructions so it runs automatically. Her block uses "SHOULD" and "AVOID" rules — clear simple language, spartan and informative, short impactful sentences, active voice, "you/your," practical insights, data and examples — paired with hard avoids: no em dashes, no "not just this, but also this," no metaphors or clichés, no generalizations, no "in conclusion" setups, no hashtags, no semicolons, no markdown, no asterisks. Then a long banned-words list (delve, embark, realm, game-changer, unlock, tapestry, pivotal, intricate, furthermore, harness, groundbreaking, cutting-edge, testament, landscape, and dozens more). Her own observed result after appending it: more direct, less hype, "no markdown, and I only counted 1 em dash." Crucially, she frames it as step one: "remove obvious AI telltale signs and start developing your own unique style" by appending your samples.

Ruben Hassid (LinkedIn AI creator, co-founder of EasyGen) is the man behind the viral "go to Wikipedia, search 'Signs of AI writing,' copy the whole page into a file" method. His recap of the Wikipedia guide is itself a clean framework: (1) stop padding sentences with empty filler; (2) ban AI vocabulary; (3) vary your pacing — don't write like a metronome ("Bim. Bam. Boum."); (4) kill the meta-commentary, just say the thing; (5) write as you speak, using "I" and "you"; (6) delete "in conclusion"; (7) use formatting "like salt," don't overdo it; (8) take a stance instead of hedging; (9) ditch the rule of three; (10) be concrete. His example contrast: AI writes "This development highlights the ongoing evolution of the digital landscape and underscores the importance of adaptability," where a human writes "This changes how small teams compete. They can move faster now."

His deeper philosophy: the magic prompt does not exist; build a living voice file instead, and update it as AI patterns drift — he notes "unlock," "harness," and "leverage" are becoming tells the way "delve" already did. He also warns against overfitting: don't force jokes, don't make every sentence a punchy fragment, don't insert slang to sound human, and don't avoid a useful word just because it's on a list. "Write normally first. Then remove the parts that sound machine-made." His test for any draft: "Does this sound like something I would actually write, or does it sound like an AI trying hard to imitate me?"

A shared warning from both, and from Wikipedia: AI detectors are unreliable, so don't write to beat them. Hassid likes to point out that detectors have flagged even the Bible as overwhelmingly AI-generated. The peer-reviewed version of this caution is stronger: a 2023 Stanford study (Liang et al., Patterns) found that GPTZero and six other detectors misclassified 61.3% of human-written TOEFL essays by non-native English speakers as AI-generated. The goal is not to fool a classifier. It is writing that actually reads like a specific human.

Copy everything in the block below into ChatGPT, Claude, or Gemini. Paste it at the start of a chat, or save it into your custom instructions / project so it runs every time. It explicitly bans the documented signs of AI writing and tells the model to match your voice.

The master prompt above removes the robot. To make it sound like you, add a short voice section. Three steps:

For one-off jobs you don't want to over-engineer, you can also just take any AI draft and tell it: "Rewrite this so it reads like a specific, clear human. Remove every pattern from my rules above. Keep my ideas, POV, and length. Output only the edited text." This mirrors Hassid's "Anti-AI-Voice Editor" prompt and Sabrina Ramonov's one-line "humanize: {your text}" approach.

Start here (first 15 minutes): Copy the master prompt into ChatGPT or Claude and use it on your next real task — an email, a LinkedIn post, a memo. Compare it to an unprompted version. You will immediately see the difference.

Week one: Build your voice profile from 3–5 samples and paste it into the master prompt. Save the combined version to your tool's custom instructions or a Project so it runs automatically. If you want to go deeper, run Hassid's 100-question interview to extract your voice instead of describing it yourself.

Ongoing: Always do the 2-minute human editing pass before anything goes out under your name — read aloud, cut the first sentence, add one specific number and one genuine opinion. This single habit matters more than any prompt.

Thresholds that should change your approach: