Can AI Really Mimic Human Writing Style? My Experiment with AI-Generated Texts
A Story of how I Tested the Limits of AI-Generated Texts and What I Learned About Their Detection
Have you ever wondered if AI can mimic human writing style? As a blogger, I was curious to find out, but I wasn’t willing to sacrifice my writing integrity to find out. Instead, I decided to experiment to test the limits of AI-generated texts and their detection.
As an AI enthusiast, I was curious to explore the methods used to detect AI-generated text and find ways to avoid detection.
The beginning
My journey began with researching the most common method used to detect AI texts, which is using perplexity, a measure of how well a language model can predict a sequence of words. If the perplexity score is too low or high, it can indicate that the text was likely generated by an AI model.
I wanted to avoid detection, so I researched a technique called “mixed-initiative interaction.” This involves a back-and-forth exchange between the AI and a human, where the human provides feedback and guidance to the AI. The idea is that by doing this, the AI can learn to mimic human writing style and avoid detection by other AI models.
The Approach
To test the mixed-initiative interaction technique, I wrote prompts engineered by myself and other AI models to generate texts with a human tone, mistakes, and flow. I fed these prompts to ChatGPT and modified them based on the results of my tests.
Initially, the mixed-initiative interaction technique worked like a charm. Some websites such as writer.com, GPTzero, OpenAI’s AI text classifier and Crossplag were fooled by the AI-generated text.
However, Copyleaks were able to detect the content as AI-generated. After further testing, I found that Copyleaks was able to detect AI-generated content with an accuracy rate of 99%.
Finding a way
I was not discouraged and decided to try another technique — AI paraphrasing. I discovered some of the best AI paraphrasing tools out there, including Quillbot and Spinbot. I used these tools to paraphrase my AI-generated text, and it worked! The AI-generated text passed undetected by Copyleaks.
Conclusion
In conclusion, while AI has made significant advancements in generating human-like texts, there are still limitations and challenges to overcome. My experiment taught me the importance of maintaining writing integrity and relying on my own skills as a writer. We must continue to innovate and experiment to stay ahead of the curve. Remember, it’s AI vs. AI in the future, and the winner will be the one who can innovate and adapt the fastest.
Subscribe to my (free) weekly newsletter about writing and web development — https://sarveshh.beehiiv.com/subscribe
