AI detectors flagged the Bible.

Hey there!

In 2023, a reporter at Ars Technica pasted the US Constitution into GPTZero. The verdict came back: "Your text is likely to be written entirely by AI." ZeroGPT agreed. The same reporting found that selections from the Bible get flagged the same way.

Edward Tian, the student who built GPTZero, explained why. Texts like the Constitution "have been fed repeatedly into the training data of many large language models," so the models learned to produce writing that sounds like them.

Scripture is the most reproduced text in human history. Every model on earth trained on it. Which is how the King James Version ends up reading, to a detector, like something ChatGPT wrote.

Now think about what that means for your ministry. A board member runs your annual report through one of these tools and gets a red score back. Now there's a conversation happening about your integrity that has nothing to do with whether you wrote the thing.

The detectors are wrong often enough to be useless. The worry underneath them is real, though, because most AI drafts do sound like AI drafts, and your people can tell.

So today I'm going to show you the 6 fixes for the real problem: making AI-assisted copy read like a person wrote it, so nobody reaches for a detector in the first place.

Let's walk through each one.

1. Stop treating AI detector scores as evidence of anything.

Detectors measure predictability, then call the result AI.

The technical term is perplexity. The tool reads your sentence, guesses which word should come next, and scores how surprised it is by the word you actually used. Low surprise gets labeled machine writing.

Scripture, legal documents, liturgy, and carefully edited prose all score low. That's what good editing does. It removes the surprises.

The research isn't close:

• A Stanford-led study ran 7 popular detectors against essays by non-native English writers. Average false positive rate: 61.22%. All 7 detectors unanimously flagged 18 of the 91 essays as AI-written.
• Rewriting those same essays with fancier vocabulary dropped the false positive rate to 11.77%. Same author, same ideas, opposite verdict.
• OpenAI pulled its own detector off the market in July 2023 and said the accuracy was too low to keep offering it.

If the company that builds the models can't detect its own output, the browser tool your volunteer found can't either.

Set your policy before you need it. One sentence in your communications guidelines: this ministry evaluates content on accuracy and sourcing, and does not treat AI detection scores as evidence. Send it to your board chair this week.

2. Read Wikipedia's "Signs of AI writing" guide and edit against it.

The best catalog of AI writing tells is free, and volunteers built it by hand.

Wikipedia editors have been cleaning up AI-generated articles since 2023 through a group called WikiProject AI Cleanup. Along the way they wrote up what they kept finding, in a document called "Signs of AI writing." It runs roughly 15,000 words, and every pattern in it came from a real article somebody had to fix.

It covers language tells, structural tells, formatting tells, citation tells, and the chatbot artifacts that leak in when somebody pastes straight out of a chat window.

Read it once, all the way through. You will recognize your own drafts on almost every screen. Then keep it open the next time you edit.

That guide is the source material for everything below.

3. Install the humanizer skill from GitHub so the rules run without you.

A checklist you have to remember is a checklist you'll skip on a deadline week.

Siqi Chen handed the Wikipedia guide to Claude Code and had it build a skill that applies every rule automatically. It sits on GitHub at github.com/blader/humanizer under an MIT license, free to copy and free to modify. Dozens of forks exist now, some tuned for academic writing, some for product copy.

Setup takes about 4 minutes:

• Drop the folder into ~/.claude/skills/ if you use Claude Code or Claude Desktop.
• Or zip the folder and upload it under Settings, Capabilities, Skills on claude.ai.
• Then type "humanize this" over any draft, and the skill runs the whole sweep before it hands the text back.

I run a customized version called humanizeit on every piece of copy that leaves my desk. Mine is up to 43 patterns, and it still catches things I stopped seeing years ago.

4. Delete the 7 tells that give away almost every AI draft.

If you fix nothing else, fix these seven.

Em dashes. Models love them. Replace every one with a comma, a period, or parentheses.

The rule of three. "Clear, compelling, and consistent." Models group everything in threes to sound complete. Cut to two or stretch to four.

Contrast reversals. "This is not a strategy, it is a mindset." Once you notice this one, you will see it in every AI paragraph you read for the rest of your life. State what's true and stop.

Inflated significance. "Stands as a testament to." "Marks a pivotal moment in." Nothing in your Sunday recap is a pivotal moment. Say what happened.

Vague attribution. "Experts say." "Many believe." "Research suggests." Name the expert or cut the claim.

Present participle tails. "...highlighting the importance of community." "...reflecting our ongoing commitment to discipleship." Those clauses carry no information at all. Delete them.

Promotional filler. "Rich history." "Vibrant community." "Breathtaking views." Your church is not a brochure.

Run a find-and-replace for those seven before anything goes out. Five minutes of work, and most of the machine smell is gone.

5. Add the 3 things a model cannot invent.

Clean copy still reads hollow if there's nothing real inside it.

A model can write a grammatically perfect paragraph about your discipleship pathway. It cannot know that 14 people signed up on Sunday, that 9 of them came out of the Tuesday night group, or that Marcus asked the question that made the whole room go quiet.

Every piece you publish should carry at least one name, one number, and one moment from an actual week.

"Marcus" beats "one attendee." "37 volunteers" beats "a strong turnout." What somebody said in the parking lot beats "the response was encouraging."

Write those down before you open the AI tool, then feed them in as raw material. Now the draft has something inside it that no model could have produced.

6. Update the skill every time you catch yourself making the same correction twice.

Here's the part almost everyone misses.

A humanizer skill is a running record of your own editing decisions. Every time you rewrite the same phrase for the third time, you're looking at a rule you haven't written down yet.

My rule is simple. Same fix twice in one week, it goes into the skill file that night. That's how humanizeit grew from a generic 15-pattern list into a 43-pattern document that knows I never use the word "quietly," never say that something "compounds," and never accept a horizontal rule in a draft.

The skill gets sharper every week because I keep feeding it what I catch.

So how do you find all the corrections you've already been making, without reading back through six months of chat history by hand?

You have Claude do it for you.

Below the fold, I'm giving you the exact prompt I use to audit my own chat history, pull out every repeated correction I've made to AI drafts, and turn those corrections into new rules for the humanizer skill. It works best inside a Claude Project, where your past conversations already live together.

The correction audit prompt.

If you're not using Projects, run it in Claude Code with your chat exports in the folder.

Paste this into a Claude Project that holds your recent writing conversations. 👇🏼👇🏼👇🏼