Debunking AI text “humanizers”: why Toolsmart failed?
I experimented with Gemini, Toolsmart, and 4 AI detectors. The results will surprise you: it turns out that many detectors lie, and here are the ones
I’ve learned about a new AI that promises to transform any AI-written text into human-written text. Since I don’t take advertising at its word, I decided to conduct a small experiment involving six AIs.
Experiment
First, I ask Gemini to write an article in which horoscopes for tomorrow will be written for all zodiac signs.
Gemini wrote me a text in a couple of seconds, which I will test now.
Next, I copied this text and sent it to Toolsmart AI to transform into human language. It transforms in a split second.
In addition to translation, this AI can do a lot of other things.
After the text became human, I found the most popular AI text detectors on Google. I started with the first one, in order. The results were very surprising.
Gptzero
The first AI detector was gptzero.me.
It gave such results for the Gemini text:
100% AI-generated. I agree. Gptzero identified it correctly.
Next, I send the text, rewritten by the Toolsmart text humanizer, for review:
It also says that the text is 100% written by AI. Ok. The first AI detector shows that Toolsmart is not working.
But to be fair, I decided to check the text that I wrote myself. What will the detector say?
The detector says that what I wrote was written by a human. Good.
But to be objective, I decided to check it on another AI detector.
ZeroGPT
I’m checking the text written first, just like last time, Gemini:
ZeroGPT was ahead in that the text was written by AI, not 100%, but 98.04% hmm, a small inaccuracy. We know that it is 100%, since I copied it directly from the AI.
Next, I add the text from the humanizer:
ZeroGPT is again not quite sure, and instead of 100% it only determined 99.2%.
I wonder what it will say about text written entirely by a human?
Honestly, I’m surprised. ZeroGpt defined what I wrote as 18.02% written by AI. I never thought that I would retell my story in the language of a robot. It’s a bit offensive.
ZeroGpt, although it determined that the text written by the Toolsmart service was most likely AI, also called human text a robot. Completely inaccurate AI. Paranoid, everything seems artificial to it.
Justdone
The next AI that will take part in the check is Justdone.
So, according to tradition, I will check the text created by Gemini first:
It identified this text as written by a robot only by 70%. It works either unfairly or inaccurately. Because it immediately offers me the use of its built-in humanizer.
If it can’t cope with determining the percentage of someone else’s text written by a robot, then how well does it rework texts? I doubt that it copes with this task.
Now we check the text created by the humanizer.
According to Justdo, the text, after using the Toolsmart humanizer, became more robotic; instead of 70%, it became 83%.
I wonder if it can handle human text?
Seriously? 81% written by a robot? I’m 81% robot…
I got the impression that this detector randomly inserts numbers so that people buy a subscription and use their humanizer. Since my text for it is more robotic than the text written by Gemini.
Scribbr
And we check the latest AI detector — Scribbr.
I added the text prepared by Gemini:
He thinks that only 33% is written by a robot — incorrectly.
Now I’m adding the text written by the humanizer:
Scribbr doesn’t want to label AI, unlike the previous detector, and says that there are almost no AI words in the text, only 13%. Wrong again.
I check my text:
Here, the service coped and did not find any traces of AI in my text, but it also almost did not find them in texts written by robots. This service is also not objective.
As a result, the ratings were distributed:
The first AI detector turned out to be the most accurate. Although the essence of the experiment wasn’t to determine the best AI detector, but to see if the Toolsmart humanizer works. In my opinion, it doesn’t, because each of the detectors determined that the text was written by a robot.
If there are any humanizers of the test that really cope with this task, I haven’t met them yet. Perhaps they just haven’t been made yet, but progress doesn’t stand still.
In addition to this, you can now also see which detectors lie and how much.
My experiment showed that Toolsmart doesn’t fulfill its main task, and many detectors give inaccurate or even confusing results.
Thanks for reading to the end! If you like my content, you might also be interested in what I do outside of debunking.
Here are some of the interesting and useful things I create.
Have you tried any humanizers? How well do they do their job?
You might also be interested in my articles:
“Everything You Say, I’ll Use It Against You in Court.” — GPTChat
GPTChat is probably not used only by the lazy. It helps someone with their work, and sometimes even enables them to…
If the article was useful, like it ❤️ and share the link with your friends✌🏻