My HumanizeMyAI Journey

I'm building HumanizeMyAI, and before writing any marketing I wanted the uncomfortable answer: do these tools actually work? So I ran the same passages through 9 humanizers and scored every output on 6 public detectors (GPTZero, Originality, Copyleaks, ZeroGPT, Turnitin, QuillBot). 162 measurements.

It was worse than I expected:

• Most "humanizers" are just synonym-swappers. The harder they reword, the more they trip the detector's paraphrase-pattern signal.

• QuillBot's humanizer scores ~95% AI on QuillBot's own detector. The vendor catches its own tool.

• The paraphraser-class tools landed between ~22% and ~33% mean AI — nowhere near "undetectable."

But the deeper problem isn't the tools, it's the detectors. A Stanford study found they falsely flag non-native English writers 61% of the time vs 5% for native speakers. A lot of flagged people didn't cheat — they just don't write like the training data.

That's the actual reason I'm building this: not "beat the detector," but help the students who get wrongly flagged — using a corpus of 2,590 real student essays (58% of them ESL) instead of a black-box paraphraser. Free, 4 runs a day.

Still very early ($0 MRR, building in public). Two things I'd love takes on:

1. Is publishing my own benchmark that names competitors honestly smart positioning, or a fight a small player can't win?

2. For a free-first tool in a gray-area niche, what converts to paid without feeling scummy?

Happy to share the full detector-by-detector numbers if useful.