Hey — I'm building an AI humanizer in public, and I wanted to share what I'm learning before there's anything to launch.
The problem I'm seeing
Most "humanizers" I've tried just swap synonyms. You know the output instantly:
"Utilize" instead of "use." Perfect grammar. Zero rhythm. Every sentence the same length.
It passes a thesaurus check and fails the human check. Real people write messy — fragments, dashes, short punches after long rambles. That's what I'm trying to capture.
What I'm building (honestly, still early)
No beta yet. No launch. I'm building a small tool that rewrites AI-sounding text so it aims to sound like a person wrote it — rhythm, burstiness, tone shifts, not just word swaps.
Right now I'm just testing in the open and learning from real examples.
What surprised me so far
- Rhythm beats vocabulary. Varying sentence length does more for "sounds human" than any synonym swap.
- Imperfection is a feature. A fragment. An em dash. A blunt short sentence. Readers trust it more.
- Context matters more than cleverness. Same input needs a different rewrite for an essay vs. a tweet vs. an email.
A tiny before/after (from my own testing notes, still rough)
Before (typical AI):
"In today's fast-paced world, it is important to utilize effective tools for content creation."
What I'm aiming for:
"Everyone's drowning in AI-sounding content. I'm trying to make mine sound like I actually wrote it."
Still rough. That's the point of building in public.
What I want from you
Drop a sentence that screams AI in the comments — I'll reply with how I'd approach rewriting it, and what I'd still get wrong. Teaching in the open, learning in the open.
If you want a note when the early access list opens, follow along. No spam, just the build log.
Top comments (3)
Your approach to capturing the nuances of human writing by focusing on rhythm and imperfections is spot-on. It’s fascinating how sentence length can influence perception so significantly; I’ve noticed similar patterns in user engagement when content feels more conversational. As you continue to refine your AI humanizer, it might be worth exploring how machine learning models can incorporate user feedback loops to adapt their outputs over time. If you're looking for additional engineering support in this area, I'd be glad to explore a paid collaboration! What specific challenges do you foresee as you develop this tool further?
Thanks Luis — great question on challenges. Three I'm hitting right now:
Context window limits. The tool needs to understand the purpose of the text (essay vs tweet vs email) to rewrite appropriately. Same input, different output depending on context.
Preserving intent while changing rhythm. You can't just scramble sentence lengths — you need to keep the meaning intact while making it feel human.
Measuring 'humanness.' There's no clean metric. Perplexity scores, burstiness ratios — they help, but the real test is: would a human flag this as AI?
On feedback loops — absolutely. The ideal loop is: user submits AI text → tool rewrites → user rates how human it feels → that data trains the next version.
Happy to chat more about the engineering side. What's your background in ML/NLP?
Some comments may only be visible to logged-in visitors. Sign in to view all comments.