Have you ever wondered how to make AI-generated text sound more human-like? I know I have. In fact, I’ve spent months trying to create an AI humanizer, only to realize that it’s not as easy as it sounds. My first approach was to use the Llama-3.1-8B model, fine-tuning it to classify between AI-generated and human-written text. But that didn’t quite work out. I then tried a modified RL approach, fine-tuning the model to rephrase existing AI-generated text and optimizing the humanness score. Sounds good on paper, but it didn’t yield the desired results. I repeated this process several times, training new scorers similar to the GAN framework, but unfortunately, it was largely unsuccessful. Looking back, I think a T5 model would have been a better fit for this task. But I’m not alone in this quest. If you’re also interested in creating an AI humanizer, I’d love to hear from you. Do you have any suggestions, links, papers, or models that you can recommend? I’m specifically looking for open weights/open source models, not paid APIs. Perhaps together, we can crack the code and create an AI humanizer that truly sounds human. Share your thoughts and experiences in the comments below!
