Please bear with me for 5 minutes.
AI detectors are crap. They detect certain words and phrases and incorrectly mark human writing as AI and AI writing as likely human. I have tried Winston Premium and Originality Premium with similar results. As a ghostwriter who has to give the promise of not using AI, this is a huge burden for me. When I send my articles without checking (manually written, mind you), they always carry a 50% chance of coming back as being flagged for AI by some obscure, random AI detector. Clients relying on a bunch of them makes things worse.
So, lately, I’ve been experimenting with this. If I avoid certain words, it comes clean, but it’s not a hard science. My results are not consistent. Then, I wondered what if I DID use AI? I purchased a Gemini Ultra subscription for this. I supplied Ultra with my own writing, then gave it a few instructions along the lines of making the text a little casual, incorporating high burstiness, varying sentence lengths a lot, and avoiding common AI words (without specifying any such words). The results were squeaky clean.
But again, this isn’t perfect and chatbots tend to forget their instructions or get inspired by random chance over time, I don’t know how, even though Gemini is supposed to have an insane “memory”. So, then, I looked at all these open-source models. Ultimately, the topics I write on are pretty basic. Health, fitness, finance, insurance, law, etc. Even something like Llama 3 or Mixtral 7x22B will have a good idea on a topic in any of these niches.
But using them was a different experience. They performed terribly, to say the least. Detectors like Originality are almost 100% accurate in detecting low-effort AI content (As we embark on a journey…).
Then I saw so many fine-tuned instruct versions of these models. Some are good at coding, some at comparisons, some at other miscellaneous stuff. And that got me wondering, is there any model fine-tuned with custom tokens and instructions in a way that helps content writers avoid AI detection? If not, what’s the closest you can get?
Mind you, I can’t do the fine-tuning on my own because the process of tokenizing my previous 500 articles on a niche to write the next topics, let’s say, is going to be more time-consuming than writing the next topics myself, using AI detector, and rewriting the parts being flagged.