A new way of making ChatGPT and Claude sound more human.

Aug '23 in Writing

I've been running multiple experiments trying to figure out how to get these 2 AI language models to write more human-like and not get detected as AI generated. Then it hit me, there are famous people in my profession who share my target audience. These individuals are wildly successful at targeting the audience that I want to target. They are authors, they appear on television, and they have hundreds of successful articles considered to be the authority in the field. Why not ask AI to help me find these people, describe their writing styles, and then I can instruct AI to copy them?

I won't make this post too long but I will share the idea and hopefully you can improve this. I asked AI to provide me the top 10 most prominent medical correspondents on popular tv news programs. These people are successful at targeting my audience. Then I asked AI to provide short paragraphs highlighting each of their writing styles. I picked 2 contrasting styles that I liked. Within my extensive prompts, I asked AI to write in the styles of both of these correspondents.

In my experiment, the results were fascinating. The tones flowed between these 2 authors very well. The sentence structures varied and the emphases changed appropriately to fit the context of the paragraphs. The results were the best of both authors. I've been able to evade Winston AI detection and passed all detection on Undetectable.AI. It might be possible to use 3 authors or more. This should make the articles even more human-like as the style, perplexity and burstiness will be high. You could also ask AI to describe the author's style, modify the description of their style and add that to your prompts.

So far, this has been a game-changer for me in my experiment and I plan to start producing several blogs this week using this newly discovered technique. You can identify prominent people in your field or your area of interest and run this experiment too. And of course, PLEASE share any improvements you find.

19 comments