You didn’t really “do something wrong.” You ran into the hard limits of what these so‑called humanizers can actually do.
Couple of things to add on top of what @mikeappsreviewer and @byteguru already tested:
-
Detectors and humanizers are playing different games
Originality’s business is detection. Their incentive is to keep that side strong. The free humanizer is, at best, a side feature. Expecting it to reliably fool their own detector is kind of like expecting an antivirus vendor to ship a free virus creator that always slips past their scanner. Even if they tried to make it good, they have to be careful not to undermine their flagship product. -
“Mixed results” is normal, not a bug
Those “partly human / partly AI” scores show how noisy detectors really are. Same text, slightly edited, can swing the score a lot. That is why your mixed outcomes do not say much about the quality of the humanizer. They mostly say the classifier is sensitive to small surface changes but still tracking the same underlying patterns. -
Light paraphrasing will almost always fail
Humanizers that only:
- replace some synonyms
- add some filler sentences
- keep paragraph structure intact
are basically lipstick on a pig. Detectors look at: - burstiness and uniformity
- how predictable the next token is
- repetitive phrasing and “AI cadence”
If structure and rhythm stay the same, detection usually stays the same too. That is why you got those “100 percent AI” outputs even after using the tool.
- Where I slightly disagree with the others
I don’t think the solution is always “aggressively reshape the text.” If you go too hard, you get weird, over‑noisy writing that also looks fake in a different way. Detectors might pass it, but any human editor or teacher will raise an eyebrow. The better play is:
- moderate automated rewriting
plus - small, very human edits that tools are bad at: abrupt transitions, personal asides, oddly specific details, mild contradictions, formatting quirks.
- Reliability of humanizers in general
As of now:
- They are not reliable enough to bet anything serious on them.
- They can reduce the AI score sometimes, but not predictably.
- New detector update, and your “bypass” text might get nailed tomorrow.
- If you still want to use tools
Since you mentioned Originality specifically, a realistic stack looks more like:
- Use something stronger than a light paraphraser, like Clever Ai Humanizer, which actually tries to change structure and style more aggressively.
- Then manually edit the result. Add your own perspective, change ordering, cut whole sections, and inject your real voice.
The key point: treat Clever Ai Humanizer or any similar tool as a drafting or reshaping step, not as a fire‑and‑forget “make this pass Originality” button.
- The harsh truth nobody selling tools will say out loud
If the goal is “pass every detector all the time using pure AI,” you are going to keep chasing your tail. Detectors improve, humanizers adapt, repeat. The only stable way to look human is to actually write like a human, with AI as a helper instead of a ghostwriter.
So no, you didn’t mess up. The marketing around “AI humanizers” is just overselling what is mathematically possible right now.