Introduction
The rising popularity of AI Text Generation has resulted in a panic among those whose career was built upon the creation and evaluation of text.
To these belong, journalists, data analysts, teachers, and professors. While the two former can only express their misery about such developments,
the two latter may believe that they have found a solution. Just around the time AI Text Generation became known to the broader population, AI text detection software has been introduced.
There are many of them, the most popular is called GPTZero.
There is however the problem that, because most educators, and the public at large, do not understand how AI text generation and AI text detection work.
They believe the detector to magically be able to differentiate between human and AI text, which is a dangerous assumption to make, at least for the students who will suffer from that
non-understanding.
How Detectors works
AI text detection works by evaluating two aspects of a text; - the word variety (Perplexity) and - the sentence lengths (Burstiness). Each human is different. Each one of us grew up in a different environment, learned different things, and prefers different styles of expression. That is reflected in how one writes as well. When a human writes, he maybe remembers cool words that he wants to use. The words used by a human vary in difficulty. An AI however has no such restrictions. An AI has access to the entire dictionary at once and writes whatever word is best to use. A text-generating AI produces text by predicting the most likely word to follow after the one it wrote before. Since it has access the basically every word, an AI-produced text has a very low variance in word choice. This sounds like a good idea until one realizes that a human writer might not have a high word variance either. If you are either generally well educated or simply decide to keep a constant style throughout the text you could end up with a low word variance and therefore tripping up the AI Text detector. The second way the AI-Detectors recognize AI-generated text is via the sentence lengths. Similarly, as people have learned differently and have their own styles, their sentence structures will vary as well. AIs tend to produce very similar sentences and will not often vary their lengths throughout a paragraph. A human does that quite often, for better or for worse.
Examples and Bypassing
If you are interested in if you are possibly writing like an AI, you can try it out for yourself. ZeroGPT has, as of the writing of this article, no word limit, unlike many other services. I tried it for myself and even texts that I have written before the “great AI revolution” are still being marked as at least containing AI-generated text. The exact percentage may vary, but it is clear that using such detectors as a reliable way of checking for AI text is useless. This is also due to another factor. You can trick the detectors. There are a multitude of strategies to trick these detectors that require more or less work on your part. Next to AI Bypassers (like Netus AI for example) you can either paraphrase or retranslate a text. When retranslating, you generate text in a different language and then use DeepL to translate it into English. Alternatively, you can also generate an English text and translate that into a different language and then back. Paraphrasing is the most secure way of bypassing detectors, while also being the method with the most manual labor involved. Generate a text, read through a paragraph of it, and write it down using other words. Further examples of AI Detectors are:
And here are some AI Bypassers:
In Conclusion
While AI Text Detectors can be used to detect the wrongdoings of the laziest of students, as soon as anyone puts thought into disguising their usage of such generators detection becomes difficult and unreliable. Simultaneously, even if you do not use AI to generate your text, you might accidentally write like one and therefore have to check your text for it, in case a professor or teacher might suspect you of using it. One has to accept that AI Text generation will stick around and at the very least fulfill the same role that calculators do in mathematics.
Technology