How to Humanize AI Drafts for Your Newsletter

Newsletters are the most voice-sensitive content on the internet. Subscribers paid you (with their attention, or their money) because they wanted to hear from you. The moment an issue lands in their inbox and reads like a SEO blog post written by a content farm, they stop opening. Open rates do not lie. A 42% open rate that drifts to 28% over two months is almost always a voice problem, not a deliverability problem.

At the same time, you are running a business, a job, a life, and a newsletter at 11pm on a Tuesday. So you use AI. There is nothing wrong with that. The problem is the workflow most people use, which is: prompt ChatGPT, paste the output, hit send. That output has a cadence, a vocabulary, and a structure that subscribers learn to recognize within two issues. This guide is the workflow that gets you the speed of AI drafting without the recognizable AI output.

Why subscribers can tell, even when they cannot say why

Most readers cannot articulate why a piece of writing feels off. They will not write back saying "your sentence length variance has collapsed and you are using too many transition adverbs." They will just stop opening. The signals are real even when the diagnosis is not.

AI-drafted newsletters tend to share a handful of patterns that build up like fingerprints across an issue:

• Uniform sentence length. Most sentences fall between 14 and 22 words. Real writing has three-word sentences next to forty-word sentences.
• Transition adverbs at the start of paragraphs. Furthermore, Moreover, Additionally, In conclusion. Real writers almost never use these in casual writing.
• Lists of three. Every claim gets backed by exactly three reasons. Every recommendation has three steps. The pattern repeats across sections.
• A summary paragraph that restates everything. AI drafts always want to recap. Newsletters almost never need this.
• Vocabulary that sounds like a TED talk. Words like "navigate," "leverage," "harness," "embrace," "unlock," and "elevate" appear two or three times per issue.

None of these are wrong individually. Stacked together, across a whole issue, they read as machine output even if no detector is involved. Your subscriber's brain is the detector. It works on vibe.

The workflow: seven steps from prompt to send

This is the workflow we recommend to newsletter writers who use AI for drafting. It is designed for issues that take you under an hour to ship, not for tentpole pieces that take a week. The whole point is to preserve speed while restoring voice.

1. Do not ask AI for the whole issue

The biggest mistake is asking AI for "a newsletter about X." You get back a complete-feeling artifact that you then have to undo, which is harder than starting from scratch. Instead, ask AI for the parts that are mechanical: a research summary, a list of relevant examples, a draft of the middle section that explains a concept. Write the parts that need voice yourself: the opener, the closer, the transitions between sections.

Practically, this means your prompts should sound like a brief to a research assistant, not a brief to a ghostwriter. "Summarize the last three weeks of news on X in bullet points" is a good prompt. "Write a 700-word newsletter section about X in my voice" is a bad prompt, because AI cannot write in your voice, and asking it to try produces a strange smoothed-over imitation that reads worse than straight AI output.

2. Kill the SEO-blog cadence

Default AI output reads like a SEO blog post. It opens with a problem statement, introduces the topic, gives three subheadings, and closes with a takeaway. That is not how newsletters work. Newsletters open with a hook (a story, a sharp opinion, a personal admission) and end with a turn (a question, a recommendation, a link to next week).

Before (AI draft): "Building a strong content strategy is essential for any modern business. In this issue, we will explore three key principles that drive content marketing success."

After (newsletter): "I rewrote our content strategy last weekend. Threw out about 80% of what we had. Here is what I kept, and why."

The "after" version commits to a specific event, names a number, and promises a reason. The "before" version says nothing. Every opener in your newsletter should be the second kind.

3. Restore your hooks

Every newsletter has a few signature moves. Maybe you always open with a one-line story. Maybe you use a recurring sign-off. Maybe your section breaks have a specific format. These are your hooks and they are how subscribers recognize you. AI drafts strip them all out because they are weird, specific, and not "best practices."

Make a one-page reference doc with your hooks. Three to seven items max. Examples:

• "I always open with a sentence about what I did this week."
• "Section headers are questions, not statements."
• "Each issue closes with a specific recommendation to try."
• "I sign off with the same three words."

After the AI draft is in place, walk through the document and add back each hook by hand. This takes three minutes and is the single highest-leverage step for keeping your voice.

4. Vary sentence length to your speaking rhythm

AI sentences run 14 to 22 words almost uniformly. Your sentences do not, because you are a person and you speak. Read the AI draft out loud. Anywhere you stumble, the sentence is too long. Anywhere it feels flat, two sentences should be combined into one longer breath. Anywhere it feels formal, replace the AI sentence with something you would actually say.

A useful trick: after a long sentence (over 25 words), force a short one (under 8). The contrast creates rhythm. The brain registers rhythm as "this was written by a human" without your reader knowing why.

5. Insert your specific examples

AI cannot tell your stories. That is your only durable advantage. Every claim in the AI draft should be replaced or supported by a specific example from your life or work. Not "many businesses have seen this work," but "I tried this at our last company in 2023 and the open rate moved from 31% to 44% in six weeks."

Specificity has a side effect that detectors and readers both respond to: it is hard to fake. Round numbers like "50%" feel made up. Numbers like "44%" feel measured. Even when the underlying story is the same, the specific version reads as human and the generic version reads as AI.

6. Run a single clean humanizer pass

Once you have done the steps above, the draft still has residual AI tells: a few "leverage" verbs, some uniform sentence lengths in sections you did not edit heavily, the occasional list of three. This is what a humanizer is actually for. One clean pass through a humanizer takes care of the residual patterns without flattening the voice you just added back.

Important caveat: do not run the whole issue through a humanizer before doing the voice work. If you do, you get a generic-human baseline that is still not yours, and the voice work has nothing to attach to. The order matters. Voice first, then humanize.

Tools like Metric37 are designed for this single-pass use case. Paste the post-edit draft, get back a cleaner version, do one final read-through, send.

7. The audible test

Before hitting send, read the issue out loud. Or paste it into a text-to-speech reader and listen on 1.5x. Two things to listen for: (1) anywhere you cringe or trip, the sentence needs work, and (2) anywhere it sounds like a press release, the language is too formal. This is the test that catches what every other step misses.

The Substack-specific adjustments

Substack subscribers are unusually voice-sensitive. The platform explicitly markets itself as the home of personality-driven writing. AI tells stick out more here than on, say, a Beehiiv marketing-focused newsletter.

Adjustments that work specifically for Substack:

• Open with first person within the first sentence. Substack readers expect "I" before "you" before "we."
• Use comments early. If a subscriber commented on last week's issue, quote them in this week's. AI cannot do this. It is immediate proof of voice.
• Title in lowercase or sentence case, never title case. Title case reads as a blog post. Lowercase reads as a note from a person.
• End with a question that invites reply, not a CTA that asks for shares. Substack rewards reply rates more than share rates.

The Beehiiv-specific adjustments

Beehiiv newsletters skew more toward business and creator-economy topics. Subscribers are more tolerant of structure and less tolerant of meandering. But they still detect AI cadence quickly.

Adjustments that work for Beehiiv:

• Keep AI structure (intro, three sections, takeaway) but rewrite all transitions in your own words.
• Use specific dollar amounts, percentages, and dates everywhere you can. Generic claims kill credibility faster on Beehiiv than almost anywhere else.
• Open the issue with a number or a date in the first eight words. "Our newsletter passed 14,000 subs on Sunday" beats "I have some news to share about our newsletter."
• Cut the "let me know what you think" closer. Replace with "I'm testing X next week, will report back." Specificity again.

Common newsletter mistakes after AI drafting

Even with the workflow above, certain patterns slip through because they are easy to miss in your own writing.

The "let's explore" opener

AI loves "Let's dive into," "In this issue, we'll explore," "Today, I want to share." Search and destroy. Replace with a sentence that does the thing instead of announcing the thing.

The empty intensifier

"Truly," "really," "incredibly," "absolutely," "completely." When AI does not know how strong a claim is, it hedges by stacking intensifiers. Cut them. The sentence almost always reads better without them.

The over-explanation

AI does not trust the reader to follow an argument, so it spells out every transition. "What this means is..." "The reason for this is..." "To put it another way..." Readers of your newsletter are sharper than the average web reader. Trust them. Delete the connective tissue and see if the argument still holds. It usually does.

The closing call to action that does not match the issue

AI ends every piece with a generic ask: "Subscribe for more," "Share with a friend," "Let me know your thoughts." Real newsletters close with a turn that connects to the specific issue. If you wrote about pricing this week, close with "Hit reply and tell me what your last price increase looked like." Specific closers get specific replies. Generic closers get crickets.

How long this workflow actually takes

For a 1,200-word weekly issue, the full workflow runs about 45 minutes once you have done it three or four times. The breakdown:

• Brief and prompt AI for the mechanical parts: 5 minutes.
• AI drafting time: 2 minutes.
• Restore hooks and openers: 10 minutes.
• Sentence length and rhythm pass: 10 minutes.
• Insert specific examples: 10 minutes.
• One humanizer pass: 2 minutes.
• Read-aloud final check: 5 minutes.

That is shorter than writing from scratch, and the output is recognizably yours. Compare to the alternative workflow (prompt, paste, send), which is 7 minutes but produces an issue your subscribers will eventually stop opening. The 45-minute version compounds; the 7-minute version costs you the list.

The real goal

The goal is not to fool your subscribers into thinking you did not use AI. Half of them know you do, and they do not care. The goal is to deliver a piece of writing that sounds like the person they signed up to hear from. AI drafting is a tool. Your voice is the product. The workflow above is how you keep one without losing the other.

Try it on your next issue. Track open rates over the following six weeks. If voice was the problem, the curve bends back up. If it was not, you have ruled out one of the harder variables to test, and that is also a win.