Journal
On undetectable writing
Field notes from building an AI humanizer, written by Gabriel Kilonzi. Everything here comes out of the same work: reading how detection tools score text, testing rewrites against them, and figuring out which changes actually move a score rather than just sounding like they should.
What you will find here
- Detector teardowns — what tools like Turnitin, GPTZero and Originality.ai are actually measuring, and why perfectly good human writing sometimes fails them.
- Rewriting technique — sentence rhythm, burstiness, the tell-tale tricolon, and the structural habits that mark a draft as machine-made.
- Test results — what changed in our own eval runs when we changed the engine, including the attempts that made scores worse.
New here? Start with the Turnitin guide for academic work, the use cases page to pick a mode, or how the engine works for the method behind all of it.
Jun 13, 2026
How to Make AI Writing Sound Human: 9 Edits That Actually Work
AI models write to a statistical average. They favor safe word choices, uniform sentence length, and predictable structure, which is exactly what makes the output feel flat.
Read article →
Why we write about AI detection
Most advice about "beating AI detectors" is guesswork passed around in forums: swap a few words, add a typo, run it through a thesaurus. In our own testing none of that moves a score in any reliable way. What does move a score is structure — how long your sentences are relative to each other, whether every paragraph opens the same way, whether each point arrives in a tidy group of three. Detectors measure predictability, and machine drafts are predictable in ways most writers never notice in their own work.
So the articles here start from measurements rather than opinions. When we claim a change helps, it is because we ran the same passage through the same detectors before and after and watched the number move. When a change did nothing, we say that too. Writing about the failures turns out to be more useful than writing about the wins, because it saves you from spending an afternoon on an edit that was never going to matter.
Who these posts are for
Students submitting coursework through Turnitin, freelancers whose clients run everything past Originality.ai, marketers worried about how search engines treat machine-written pages, and anyone who drafts with AI and wants the finished piece to read like their own voice. You do not need a technical background. Where a post gets into the mechanics of how a classifier works, there is a plain-language summary at the top.
A note on honesty
No tool, ours included, can promise a specific detector score. Detectors change their models without notice, and two passages of identical quality can score differently for reasons nobody outside the vendor can explain. What we can do is show our method, publish what we measured, and let you judge. If a post here ever overstates a result, tell us and we will correct it.