LARRUZ demonstrates how to bypass AI content detectors using ChatGPT. He tests a 300-word AI article against Writer Data AI, Crossplag, and Content at Scale, then uses a prompt asking ChatGPT to increase perplexity and burstiness to rewrite it, showing the rewritten version scores as human-written.
There are a number of services and systems that claim to have the ability to differentiate between human and AI generated content. However, it's very easy to bypass these systems by carefully engineering your prompts. This is exactly we are going to do in this video. We will do this within ChatGPT without the need for any external tools or websites. We will test this on four different AI detector systems. Show you an example. I asked ChatGPT to write a 300 word article on mental health. So I will show you a number of AI detection systems and see if they can able, they are able to actually detect this or not. So the first one that we're going to be trying is Writer Data AI. So let's see what it says. So According to this AI content detection system, it says that it's around 80, 58% human generated content. The second one is going to cross flag. So let's see what it says. Okay, this takes some time so that run this. It says this text is mainly written by an AI. Okay. So it's like 100% confidence. And this score shows like how confident the AI is. For character date app, the confidence level is 99.98% fake. So it thinks that it's an AI generated text. The last one that we will check is called content at scale, not AI. So let's see its prediction. Okay, so it says human content score, it's 0%, obviously AI. So it seems like all of them were able to correctly identify that this text was generated by an AI. Now let's see if we can bypass these AI contact detectors or not. Right. So for this we have a simple prompt that we will use. So let's go with the prompt. First we are saying imagine you are a content generator. So we are assigning a role to the AI system. Right. Then to write effective content, both perplexity and burstiness are important. So we are introducing two concepts that we want the AI to consider and pay attention to. Moving on. Perplexity assesses time complexity and burstiness evaluates sentence variation. So perplexity basically defines how complex the text is. Right. And burstiness defines how much variation you can have in the text. Like you, you can have shorter or longer sentences. So people often write a mix of long and short sentences while machine generated sentences tends to be uniform. Right. So people usually can have variation in the sense the length of the sentences that that. Right. But machine usually keep them uniform. So that's basically the context that we have provided to the chat GPT AI. Then the instructions are I'll provide you with an article and ask you to rewrite it. When composing the content need to ensure a suitable balance of both perplex, perplexity and burstiness. Right. So this prompt has three components. First is the role that we want the AI to assume. Then we define the context and the cons. We introduce it to different concepts. Right. And at the end what action it's supposed to perform. So that's going to be rewriting articles for us. So let's run this. Okay, so in the output it simply generated this, that. In content creation, it is crucial to strike a balance between perplexity and robust burstiness. Perplexity gauges the intricacy of text while burstiness assesses the diversity of sensors. A good mix of short and long sentences crucial in writing. Whereas machine generated text often lack this variety. So let's say we are going to give it the article. We'll say rewrite this article. Okay. And let me give it in the quotes so that it pays attention to the article itself. And let's run this. Okay, so it finished writing the article or basically rewriting the article. So let's simply copy this and go to our first AI detector. Let's run this one. So now the score from change from 58 to 98 human generated content. Right. Let's try it with the second one. Okay, while this is running, let's go to character and see what the score is. So now it's less than 1% fake. So it means that it's human written here again like 9%. So it says the text is mainly written by human. And the last AR detector system is our content at scale. So let's paste it. Right, let's see. Okay, so it's 91% human content score. So. So this looks good. So based on this example, you see that just playing around. With a carefully curated prompt which simply changed the perplexity and bursting of the text, we were able to bypass the AI content detection systems. This is a simple example. We tested like four different systems. There might be some other ones where you might have to play around with the parameters to confuse them. I hope you like the content. Please consider subscribing. Thanks for watching.
