Speaker 1: Hey guys and welcome to Words at Scale. So this is one of those videos that takes like two hours to make and this is likely to be like five minutes long and we are going to be talking about AI detection and torrenting. It still surprises me to this day how many people are concerned about torrenting and AI detection even though I have stated my position virtually in every video that I made and the position is as follows you should not be worried about AI detection as long as you are not in a YMYL niche or in academia or are not writing for your clients who want to test your outputs with AI detection so if you do not belong to any of those groups you shouldn't be worried about AI detection at all because there are lots of evidence already on the internet that AI-generated content ranks on the first page of Google and I can attest to that with my websites also. Okay, so in this video I'll be showing you a trick to check your outputs with torrenting without actually ever touching torrenting and all this sounds crazy but bear with me and also for those of you who are not interested in torrenting we'll be comparing different AI detection tools against accuracy between one another so that might be of interest to you also. And before I show you the results, let's just go over the whole process. So step number one was going into ChatGPD and generating a list of 10 most popular topics for essays. As you can see, they're numbered 1 through 10. Then I generated a 1000 word essay for each and every one of those topics. the actual word count ended up being between 500 and 600 words. And then I have chosen quite a few AI detectors and actually on my blog I have this article that I wrote a while ago which is my review of the best AI content detectors and there are quite a few and then I also tested them against three outputs the low quality AI output, the output written by a human, and a high quality AI output. So if you are interested in AI detection and you want to get an overview, you can go to TrickMeNot and the Slack is which is the best AI content detector. But the reason I'm telling you this is because I needed to select the AI detection tools for this experiment and I wanted to have something that is mostly free and without limitations. So I have chosen content at scale, which you can use without any limits. I have also chosen copy leaks. That has some limits, but they are easily bypassable, whether you switch in a browser or turn on and off your VPN. And also the cross-plaque that also has like three trials per session. But again, you can change your browser, switch on and off your VPN, and you can use cross-plaque all over again. So these three are the free versions and the originality.ai, I needed to check that as well. So I ended up with cross-plug, copy leaks, content at scale and originality. And then, of course, I have the torrenting ready. And as you can see, I have tested 20 essays. essays, so the 10 original ones that I wrote with ChatGPD, and 10 altered versions that were altered via the undetectable.ai, and the setting that I used was high school, essay, and more human. So more human gives you the best chances of bypassing AI. And the whole logic behind the experiment is as follows. So you have your four AI detection tools and you have your torrenting. First what I wanted to do is to get a reading from every AI detection tool on the list for the original essay and look at the correlations. And the logic is as follows. If torrenting shows high correlation with, say, cross-plaque, then you don't need torrenting to check your AI. can rely on cross-plug for all of your future SA testing or whatever. And the same goes for the altered or transformed versions. I ran everything through cross-plug, copy leaks, continent scale, originality. And then the ones that would show the closest numbers to torrenting would be used as proxies. We are comparing different detection tools between one another. We are using the same documents and the ones that show the most similarity with torrenting are gonna be the proxies. So we won't be needing torrenting at all, we can rely on these proxies. And this is how the experiment started. So now finally to the results. And the results are interesting. So first things first. Originality, even though it scored everything like 100% AI, it's not a reliable metric to use in conjunction with torrenting, because with torrenting you can see that SAE number 1 through SAE 5 were scored like 100% AI, but then there were the outliers like SAE 6 scoring 3, SAE 7 scoring 69 and SAE 9 scoring 44. So if you run the correlation analysis between originality version 1.4 which is the most recent one and torrenting you can see that they have a correlation of almost 40 percent which is even less than a chance and normally you would want correlation coefficient of 0.6 or 0.7 for it to be reliable and the same goes for cross-plug and continent scale so the readings were all over the place and they have negative correlation so again this is not something that we can use however with copylex basic model and basic model is the model that you use for free and if you want advanced model you need to pay the readings although none of them had 100 percent AI they correlate 75 percent of the time if you were to draw a graph let me just quickly do that so by drawing a graph we can see a visual correlation and yeah so we can see that these two lines the orange being torrenting and the blue being copyleaks they have somewhat similar peaks and dips so this means that every time torrenting thought it was AI so did copyleaks and the opposite is true every time copyleaks thought it was human torrenting did as well. So the main key takeaway here is that if you have your own methods of bypassing AI detection, if you're not using undetectable.ai for example, or if you're using prompts and on my website which is trickme.ai you still have some very good prompts for bypassing AI, you can safely or somewhat safely use copyleaks and if you see that you're getting very good scores with copyleaks Then you can assume that the same will be true for Turingtian. So this is the first step. Now I had to see how Turingtian and other AI detection tools would react to transformed or altered output, which is again altered through undetectable.ai, and as you can see still the toughest trick was originality, even though we were able to get some very high human scores, like 99% human, 94% human. The easiest, and here we cannot run correlation analysis because every time I used undetectable.ai, Torrentian returned 0% AI. So for all the iterations, for all the essays that have been transformed, Torrentian said it was 0% AI and 100% human. So there wasn't even one variation, and this is one of the reasons the correlation would not work in this example but we can look at the averages so the average score for the transformed output was zero so what's the score for cross-plaque and so what's the score for content at scale so this is in reverse cross-plaque says this is 0% AI and content at scale says it's 98% human so basically the same thing so if you are transforming your outputs and you want to check if this passes torrenting you can actually look at continent scale or cross-plag and if cross-plag on continent scale say this is zero or close to 100 when we are talking about continent scale again by using the proxy rule you can safely assume that torrenting will return the same results and basically this is it so again to sum up if you have your own ways of transforming the outputs and you just want to check the basic state of your AI-generated text, you can use Copyleaks. However, if your text has been through some sort of transformation and you don't have access to torrenting but you still want to get a feeling of how it would rank with torrenting, you can use Content & Scale or Crossplug. Content & Scale is 100% free all of the time and Crossplug is like three tests per session but again it's extremely easy to bypass just by changing your browser your vpn or basically your ip address and so on so forth so hope this video is helpful it took me a while to come up with these measurements like share subscribe if you haven't already and i'll see you next time