The arrival of artificial intelligence in society means we can’t always believe what we see. Because of this, AI detectors have recently been developed to help figure out if a text was written by one of them or by a human. In this article, we’ll look at different AI text detectors and compare them to find out which one is the best.
IMPORTANT NOTE: Even though we’ll see different AI detectors in this article, there is NO foolproof way to know for sure if something was made by artificial intelligence or a human today.
To compare the different detectors, we’ll test each one with three different texts: one written by ChatGPT-3.5, as an answer to the prompt “Act like a History teacher, write a text naming and explaining the social classes of al-Ándalus. The text should have a maximum of 240 words”; another “hidden” by paraphraser.io, which tries to make the text look like it wasn’t written by AI (we’ll see that in many cases, it doesn’t work); and another one written by a human.
Main Artificial Intelligence Detectors
copyleaks.com
When you go to this website, there’s a small problem at first: you need to sign up to have your text analyzed. However, this detector doesn’t have a minimum or maximum number of characters the text needs to have for the detection, which is definitely helpful. Let’s try the AI content detector, which is in the main menu of the website. It will tell us if the text we put in was generated by artificial intelligence or by a human, and the percentage of probability if we move our mouse over it.
Results
- Text generated by ChatGPT-3.5: It is AI content (99.9%) 🟩 - Text generated by ChatGPT-3.5 and paraphrased by Paraphraser.io: It is AI content (99.9%) 🟩 - Text written by a human: It is human text (94.1%) 🟩
Conclusions
It seems this detector is quite reliable, as it correctly identified the author in all three tests we did. It has one main advantage and one main disadvantage: the advantage is that it can detect text of any length; the disadvantage is that you have to log in.
smodin.io
We move on to the next detector, which says it can work in many languages. You don’t need to sign up to use it, but you have a limited number of uses. It will tell you how likely it is that the text was written by an AI and why it thinks so, pointing out the different sentences that make it believe that.
Results
- Text generated by ChatGPT-3.5: It is possible that the content is human and AI (50.2%) 🟧 - Text generated by ChatGPT-3.5 and paraphrased by Paraphraser.io: It is possible that the content is human and AI (50.2%) 🟧 - Text written by a human: It is likely that the content is written by humans (10.5% AI) 🟩
Conclusions
The detection isn’t very good, as it can’t reliably tell which texts were generated by AI. If you add to that the pretty simple limit on weekly uses, it’s not a very recommended option at all.
corrector.app
The next detector we are going to try is called corrector.app. After you put in the text, it will tell you the percentage of probability that it is fake. Although you don’t need to register, the main problem is that it has a maximum number of words per text: 800. Also, something quite annoying is that you have to wait 60 seconds after doing a detection.
Results
- Text generated by ChatGPT-3.5: 0.02% fake, that is, 99.98% human 🟥 - Text generated by ChatGPT-3.5 and paraphrased by Paraphraser.io: 0.02% fake, that is, 99.98% human 🟥 - Text written by a human: 0.02% fake, that is, 99.98% human 🟩
Conclusions
As we have seen, this “AI detector” is not at all reliable, at least in Spanish, even though it says it is ready to detect texts in Spanish, as its title indicates. It gives us the same answer in all three cases: the text, most likely, was written by a human being.
ai-detector.compilatio.net
We move on to the fourth detector we will see, in this case from Compilatio. This one also says it can work in many languages, as it tells us that we can write the text in English, French, Italian, or Spanish. The downside is the character limit: our text has to be between 200 and 2000 characters. You also don’t need to register, and after you put in the text and click “Start,” it will tell you who wrote the text: human or AI, and the probability of that being the case.
Results
- Text generated by ChatGPT-3.5: AI (99%) 🟩 - Text generated by ChatGPT-3.5 and paraphrased by Paraphraser.io: AI (99%) 🟩 - Text written by a human: AI (98%) 🟥
Conclusions
The results of this detector are, without a doubt, curious, as it is the first detector we have analyzed that categorizes a text written by a human as being written by AI, which makes this detector a completely unreliable tool.
ChatGPT
Yes, the fifth AI detector is ChatGPT itself, as it is supposedly able to detect texts written by it or that are likely to be written by AI, or so many people believe. Could it be true? Let’s find out below.
To do this, we will enter the following prompt, changing
By doing the tests, we get the results you can see below.
Results
- Text generated by ChatGPT-3.5: It seems to have been written by a human being 🟥 - Text generated by ChatGPT-3.5 and paraphrased by Paraphraser.io: It seems to have been written by a human being 🟥 - Text written by a human: It seems to have been written by a human being 🟩
Conclusions
I don’t know what ChatGPT will say, but it seems we have busted the myth: ChatGPT is not a good AI detector. As shown by the terrible results, we also rule out this tool.
Final Conclusions: Which is the Best AI Detector?
As we have seen throughout the article, there is only one truly reliable tool as an AI detector; the rest have not been able to correctly identify the author of all the texts. This tool is copyleaks.com, which despite requiring login to use, has been, by far, the best artificial intelligence detector in Spanish.
