Last tested October 20, 2026
I think people should know how accurate a detector is before they trust it, so I am putting all of my test results here. If something scored 40% or higher, I counted it as "caught."
To test this, I wrote a paragraph that sounds like ChatGPT and put it through the free version of each humanizer. Then I checked what came out with Optimal. Part of the score comes from an AI review, which can give a slightly different answer each time, so I ran every sample more than once. The table shows the range of scores I got.
| Humanizer | Score | Result |
|---|---|---|
| ZeroGPT | 85–95% | Caught every time. Missing words and commas, plus a hyphen character you can't type on a keyboard |
| SuperHumanizer | 95% | Caught every time. Some phrases didn't make sense and the last sentence was cut off |
| Clever AI Humanizer | 75–85% | Caught every time. Weird word choices, and some ideas were different from the original |
| Grammarly (3 full essays) | 75–95% | Caught every time. It mixed British and American spelling and used "carry out" over and over |
| Humanize AI (humanizeai.pro) | 86–100% | Caught every time. It used six different stiff phrases, like "Firstly," "It should be noted that," and "but are not limited to" |
I also tested a basic free humanizer on two short paragraphs. That one adds small grammar mistakes and splits sentences into pieces so the writing looks like a younger student wrote it. Optimal scored both paragraphs at 85% or higher every time.
Optimal does not flag plain AI writing on its own. It used to give a higher score to any text that read like an AI essay, but that also flagged real student essays, and Optimal is meant to find humanized text, not every kind of AI writing. Now the AI review's opinion only counts when the text also uses several stiff phrases that humanizers like Humanize AI and Grammarly add. Text that came straight from ChatGPT or Grok scored 0% in my tests, so a low score does not prove a person wrote something. It only means Optimal did not find signs of a humanizer.
The other half of being accurate is not flagging people who actually wrote their own work. So far I have tested four human-written essays: a student research essay about generative AI and misinformation, a short informational essay about nature, and two personal essays. I ran each one at least twice, and none of them scored higher than 15%. Before I changed how the score works, the research essay got 86% because it read like an AI essay, and the AI review also thought the nature essay was AI-written, which is the main reason I made the change. Four essays is still not enough for a fair number, so I'm still collecting more. Until then, a high score should only be a reason to look closer.
I saved every sample as a text file and wrote a script that runs them all through the same checks the website uses, including the AI review (which is Google Gemini on the free plan). Every time I change how Optimal works I run the script again to make sure the results didn't get worse. I'll update this page when the numbers change.