By Aryan Gupta · November 1, 2026
Humanizers are made to get AI writing past AI detectors, so I wanted to know if humanized writing can actually be caught. To find out, I ran 50 AI essays through a real humanizer and then tested all of them with Optimal, the humanizer checker I made.
First I had an AI write 50 essays. I made them all different kinds of writing so it would be a fair test: school essays, college essays, research reports, news articles, opinion pieces, literary analysis, business memos, history essays, science articles, and personal reflections. Then I put every single one through the free version of humanizeai.pro, which is one of the most popular humanizers right now.
Honestly, the first time I tested them, Optimal only caught 2 out of 50. The humanized essays didn't have any obvious tricks like hidden characters, and they read pretty smoothly. The AI review part of Optimal kept saying they were either a normal AI essay or just written by a person. That's exactly what humanizers are supposed to do, so it makes sense that they're hard to catch.
After that I read through all 50 myself, and I started noticing things that kept showing up in the humanized essays but not in plain AI writing or human writing.
The first thing was small mistakes that didn't match how polished the rest of the essay was. Things like "in case" being used to mean "if," or "numerous researches," or "enable to make sure."
The second thing was facts changing. In one essay "junior year" turned into "sophomore year." In another one, something that was "parallel" became "perpendicular." One essay even said the sun is 930 million miles away when the original said 93 million.
The third thing was that the writing felt flat. The essays still had the same neat structure as an AI essay, but all the fancy chatbot words were gone and the sentences were plainer and kind of stiff. A lot of them also used phrases like "the process of" and "it is hard to overestimate."
I taught Optimal to look for those things and to also count how much the writing reads like AI. I used 4 of the 50 essays as examples to teach it what humanizeai.pro's writing looks like, so it wouldn't be fair to test on those. Then I tested it on the other 46, which it had never seen before, and it caught 43 of the 46. The three it missed were all personal pieces, like a college essay and a reflection. In those, the humanizer kept the personal details from the original, so they really did read like a real student wrote them.
Usually, yes. But not by looking for just one thing. You have to look at a bunch of different signs at the same time, and personal stories are definitely the hardest. Detectors can also be wrong the other way, since some human writing sounds really generic and can look like AI. So whatever a detector says, including Optimal, I think it should be a reason to look closer and not proof that someone cheated.
All of my test results are on the Test Results page, and you can check your own writing on Optimal for free. If you want to check something by hand, I wrote a list of signs in How to Tell if an Essay Was Run Through a Humanizer.