Full disclosure: I built an editing tool for this, so I'm biased. Everything below works with a text editor, a highlighter and twenty minutes.
TL;DR
- Readers notice the same handful of patterns in machine-sounding prose: inflated significance, promotional adjectives, em-dash chains, automatic lists of three, vague attributions and a small set of overused words.
- Fix them in one pass per pattern, not by "rewriting everything". Each pass is mechanical and fast.
- Never touch your citations, numbers, quotes or claims while editing for voice. Lock them first.
- The goal is clear, specific prose that sounds like you. It is not about tricking a detector, and you should still follow your institution's rules on AI use and disclosure.
The patterns at a glance
| Pattern | What it looks like | The fix |
|---|---|---|
| Inflated significance | "plays a pivotal role in the evolving landscape of..." | Say what it actually does, with the specific result |
| Promotional language | "groundbreaking", "robust", "seamless" | Delete the adjective or replace it with a measured claim |
| Em-dash overuse | three dashes in one paragraph | Use a period, a comma or parentheses |
| Rule of three | "efficient, scalable, and robust" by reflex | Keep only the items you can support |
| Vague attribution | "experts argue", "studies show" | Name the source and cite it, or cut the sentence |
| AI vocabulary | "delve", "underscore", "tapestry", "intricate" | Use the plain word |
| Empty wrap-ups | "In conclusion, X remains a complex topic" | End on your finding or your limitation |
These categories come from Wikipedia's editor guide Signs of AI writing, which is one of the most careful public catalogues of these habits.
Step 1: Lock what must not change
Before you edit a single sentence, mark everything that carries evidence:
- every in-text citation and reference-list entry
- every number, statistic, p-value and sample size
- every direct quote
- every claim your argument depends on
Highlight them, or copy them into a side file. During the voice pass you may move a sentence around these items, but you never reword them. Most of the damage done by "humanizing" comes from a well-meant rewrite that quietly changes a hedge ("suggests" becomes "proves") or drops a citation.
Step 2: Cut inflated significance
Search for phrases like "pivotal", "crucial role", "key aspect", "in today's rapidly evolving", "landscape" and "testament to". Each one usually stands in for a specific claim you haven't written yet.
Example (made up for illustration):
Before: "Sleep plays a pivotal role in the evolving landscape of adolescent mental health research."
After: "In our sample, students who slept under six hours reported higher anxiety scores (Table 2)."
If you can't replace the phrase with something specific, the sentence probably isn't needed.
Step 3: Remove promotional words
Academic prose describes; it doesn't sell. Search for "groundbreaking", "novel" (when it isn't), "robust", "seamless", "cutting-edge", "powerful" and "comprehensive". Either delete the word or replace it with the evidence: "robust to outliers" is fine if you tested it, and then you say how.
Step 4: Fix punctuation habits
Count em dashes per page. One or two is normal; a dash in every paragraph reads like a template. Replace most of them with a period or a comma. While you're there, check for colon-heavy "Here's the thing:" constructions and rhetorical questions used as transitions.
Step 5: Break the reflex lists of three
Machine-sounding drafts love triplets: "efficient, scalable, and interpretable", "researchers, practitioners, and policymakers". For each list, ask whether you can support every item. If the third item is filler, drop it. Two honest items beat three padded ones.
Step 6: Replace vague attributions
"Scholars have long argued", "it is widely believed" and "studies show" are the most dangerous pattern in academic writing, because a reviewer will ask "which studies?" Either name and cite the source, or delete the sentence. Never add a citation you haven't read just to fill the gap.
Step 7: Swap overused vocabulary
Run a find for this short list and replace with plain words:
- delve into: investigate or examine
- underscore: show
- tapestry, intricate, multifaceted: cut, or say what the parts are
- leverage: use
- foster: encourage or support
- navigate (used abstractly): handle
Read the sentence aloud after each swap. If it still sounds stiff, shorten it.
Step 8: End on the finding, not a summary of the summary
Look at the last paragraph of each section. If it starts with "In conclusion" or "Overall" and repeats what you just said, replace it with the one thing the reader should take away, or with the limitation that matters most.
Step 9: Re-check the locked items
Compare your edited draft against the side file from Step 1. Every citation, number and quote should be identical. Every hedge should be as strong or as weak as before. This check takes five minutes and prevents the worst outcome: a fluent paragraph that now says something your data doesn't.
What I'd do today
- Pick one section (the introduction is usually the worst offender).
- Lock the evidence (Step 1).
- Run Steps 2 to 8 as quick find-and-fix passes.
- Re-check the locked items (Step 9).
- Repeat for the next section tomorrow. Editing for voice is tiring, and fresh eyes catch more.
FAQ
Is this about beating AI detectors? No. Detectors are unreliable, and gaming them isn't the point. The point is prose that a reviewer can read without wincing. If your institution requires you to disclose AI assistance, disclose it.
Will this change my meaning? Not if you lock the evidence first and re-check it at the end. Voice edits should change how a sentence reads, never what it claims.
Does it work for non-native English writers? Yes. Many of these patterns also show up in careful second-language writing, and the fixes (shorter sentences, specific verbs, named sources) make the text clearer either way.
If you'd rather not do every pass by hand
I built Academic Humanizer for students, researchers and writers whose academic drafts read machine-written. You paste your draft and get two things back: a rewrite that strips inflated symbolism, promotional language, em-dash overuse, rule-of-three padding, vague attributions and AI vocabulary while keeping your argument and citations intact, and a list of every AI tell it fixed, so you can review each change. It's based on Wikipedia's "Signs of AI writing" and costs $9.99 one time on Capafy. Still run Step 9 yourself: you're the author, and the final check is yours.
Disclosure: I built Academic Humanizer and sell it on Capafy.
Top comments (0)