AI Humanization: Make It Sound Like You Wrote It That Way on Purpose

AI humanizers are all about flying below Google's radar. Real human writing can sometimes get flagged, too.

This article originally appeared on Substack.

“Just try to make it sound like you wrote it that way on purpose,” fictional publisher Arthur Howitzer, Jr. tells inexperienced writers at his magazine office in rural France.

I love this quote so much that I’ve made it my personal motto for life—especially since so much of my work now concerns AI humanization, the art of making content created by artificial intelligence seem, well, not so artificial

Bill Murray plays Howitzer, the main character in The French Dispatch, a twee 2021 Wes Anderson movie about the French bureau of a Kansan newspaper. The Dispatch sends its writers to the town of Ennui-sur-Blasé to observe and comment. Their stories are a window into an exotic world to entertain the staid Midwesterners back home.

The movie unfurls as a series of vignettes, short stories that play upon the writers’ American perspective, the strange setting and events around them, and what their audience of tillers and toilers might want to know about the rest of the world.

As content creators, we’re now tasked with using AI tools to produce a growing mountain of words that both represent the voice and perspective of clients and topics a broadly defined audiences might want to know about. The name of the game is search engine optimization or SEO. We learn what people are searching for and make sure there are helpful words on a website that answer their questions.

But how do we make those words sound like we meant to write it that way on purpose? More importantly, how can I arrange those words in such a way that Google’s AI detector Spam Brain doesn’t catch on to what we’re doing?

The Radar Detectors of the Internet

Spam Brain, and other AI detectors watch for common patterns in writing that sound like they rolled out of a ChatGPT prompt. For AI-enabled content writers, AI detectors are like county sheriffs hiding behind billboards to catch us running a stop sign.

If Google considers you a repeat offender, it could de-index the article or even your entire site as a violation of its people-first anti-spam policy.

There’s an AI detector in Grammarly Pro, the support extension I use to flag grammar and spelling errors. According to its analysis, what I’ve written so far is 100% human-written. So far, so good!

To give us advance notice of lurking SEO cops, there is a whole raft of AI humanizers to help toss your words up in the air and make them seem more random and alive.

In my articles, I like to find sources that are not sales-y content marketing blogs. I “punch up,” looking for reputable sources—technology journalism, research papers, opinion pieces by respected thought leaders, and such. For the term “AI humanizer,” I found page after page of search results advertising dozens of products that can make your AI copy sound more like someone lovingly typed it in Arthur Howitzer’s newsroom.

Finally, among all the landing page links and “10 Best AI Humanizer” reviews, I found this article by writer Kymberli Sagita’s LinkedIn page. Sagita provides a commonsense overview of AI detection and how to evade it. She says these detectors look for “unnatural phrasing, excessive repetition, or a lack of contextual understanding.”

She recommends that writers mix up their vocabulary and syntax, lead readers through a natural chain of thought, and throw in hairy arms—flaws that send a signal that a human might be behind the curtain. She also recommends using custom models trained on the type of content you’re creating, understanding why AI detectors don’t always work, and subtly exploiting these weaknesses.

In fact, much of the corpus of human writing is also devoid of imagination and supporting details. Generic AI content is drawn from an average of corporate landing pages that don’t really say anything, vaguely worded mission statements, medical conditions directories, university program descriptions, and millions of other non-imaginative examples. Truth is, AI detectors can register false positives if the human writing wasn’t all that great to begin with.

An AI Writer in Ennui-de-Blasé

Imagine, to cut costs, Arthur Howitzer has made a decision to use ChatGPT to write all the stories for his Dispatch readers back in Liberty, Kansas.

These readers still want to know what it feels like to walk the streets of Ennui-de-Blasé, meet the colorful characters inside its crumbling buildings, and smell the Gauloises cigarettes and baking bread at the Le Sans Blague café. They loved the revolutionary spirit of Lucinda Krementz, who reported on the Chessboard Revolution while secretly helping its leader write his manifesto.

French Dispatch readers want more of that kind of writing, please.

The AI operator is an intern back in Kansas who has two choices:

1. Feed in already-published stories and fabricate the characters, facts, and events.
2. Ask Mr. Howitzer to please go out and have new experiences that can serve as the basis for the new articles.

Here’s the next problem: These colorful writers were not only describing what they did and who they met. They were also filtering it through their own memories, biased perspectives, idiosyncratic quirks, and expectations of future unfoldings.

Ultimately, the poor intern’s work will result in a drop in readership, and the French Dispatch has let its standards slip and will soon be out of business. The stories in The Paris Review are so much better.

A HUMAN Writing Framework

I’m working on a rubric for human writing. I love memorable acronyms, so I’ve organized these writing principles around the letters HUMAN. My goal is to spot issues in content I edit, whether it’s generated by an AI or written by a human being. (Humans can sound like robots sometimes, too.)

HUMAN writing is honest, useful, meaningful, authentic, and nuanced. AI writing tends to fall short in some of these categories unless it has access to original human inputs. You can quickly enhance an article using the HUMAN principles if you have an interview transcript, manuscript, notes from a meeting, podcast, or some other way to access someone’s authentic words.

According to the framework, HUMAN writing is:

1. Honest. The writing is clear, transparent, and truthful and builds trust between the writer and their audience. It does not seek to hype or hide.
2. Useful. The writer focuses on what matters most to the reader, offering answers to people’s questions. The writing is organized in a way that best enhances the reader’s experience.
3. Meaningful. The writing engages the reader with vivid, memorable detail—whether it’s personal experience, hypothetical examples, or detailed references to external sources.
4. Authentic. The writer’s voice shines through, offering perspectives, experiences, and connections that the writer is uniquely positioned to offer.
5. Nuanced. The writing is subtle, layered, and sometimes a bit off-topic. The writer trusts the reader to come along on a journey with them. Style and approach are as important as the words themselves.

HUMAN Writing supports Google’s own EEAT framework, which discovers content authority and drives search rankings. EEAT stands for experience, expertise, authoritativeness, and trustworthiness.

Human writing goes beyond the acronym to include avoiding the stochastic parrot clichés and language patterns that AI content tends to exhibit. It also includes the principles of perplexity and burstiness—two features of human writing that good AI writing mimics.

Make AI Smoothies With Human Words

I love to tell colleagues who ask about AI content generation that “I can’t make smoothies without fresh fruit.”

If I don’t have access to human memories, experiences, opinions, or connections, the writing sounds flat and ungrounded. I can run it through an AI humanizer to get rid of the C-3PO effect, but I am constantly challenged to make it sound like a real person.

I’m blessed—or maybe cursed—with the talent of hearing a voice in my head when I read things. It comes from my years of reading radio scripts out loud to get the correct timings. When I can’t hear that distinct voice because the writing literally lacks character, I find it distressing. This internal AI detector has been both my superpower and bane as more automated content keeps crossing my desk.

How are you using AI to write for you? Do you use it to process recording transcripts and provide organized notes? Does anything you write ever go public?

If you’re not trying to make your AI outputs sound like you meant to write it that way on purpose, what will the folks back in Kansas think about your talents?

This article is 100% human-written with AI editorial support by Grammarly Pro.