How To Easily Trick An AI-Text-Detector With This Simple Code

Image by Author

Currently, AI text detectors are gaining more and more popularity in various fields in order to identify content that is not entirely written by human beings.

In my opinion, this is a highly valuable step towards fighting the flood of low-quality content creation through AI.

However, since I am a data nerd, I immediately got curious how these detectors work and how they can be outsmarted. One day while playing around with transformers models on Huggingface, I suddenly had a quite different idea…

What, if I just translate AI-generated text to another language and then back to its origin? Since I have worked some time with artificial language understanding, I knew that there are indeed logical and structural differences between different languages (e.g. subject-verb-object (SVO) or verb-subject-object (VSO) following languages).

So, my hypothesis was that if I choose a language that differs much in these properties, the text could also change when translating it forth and back.

My first experience was that one language translation loop is not enough. Hence, I decided to define a set of languages in a bigger translation circle where each subsequent language differs in structural characteristics.