The Best Chat GPT Detector is Here (10 Testruns)

Ladies and gentlemen, spotting AI generated text has been the hot topic of probably January, not just within academia and schools, but also inside and throughout the wider public. People have been wondering, this AI can generate insanely good text, but how do we know? How do we know if this is written by a human or a AI? Well, up until now, the answer was kind of, well, we don't. There were these detectors that were based on GPT2, which, as you know, is old and irrelevant. And it did not work at all. I created two videos about them. In one, we completely destroyed it, right? By adding one space. The thing fell apart. By adding, in the style of Borat, it fell apart. It was not usable at all. But today, big announcement by OpenAI themselves, they released the AI text classifier. And that's exactly what we're going to explore today. Okay, so the way this thing works is you can copy paste text and it tells you, is it human, is it AI, but it doesn't tell you in clear terms, okay? It's not in percentages. It's nothing close to absolute. It tells you it's very unlikely, unlikely, unclear, possibly or likely that it is AI generated. Okay? So it's a recommendation at best. And here's a few things that you need to know before we get into this. All right? I have a few bullet points here. You can read through all the documentation, you can read through all the detailed information here yourself. But I want to give you the bullet points and they go as follows. Okay? The most important thing is if the text is below 100 to 150 to 200 words, it's not even worth trying this, right? If it's a short amount of text, I think they put it in terms of, as you can see here in the official blog article, which you can find on OpenAI.com blog, they tell you that if it's below 1,000 characters, which is roughly 150, 200 words, don't even bother like if it's a text message, like, you just don't know if it's AI or not. Okay? So secondly, if somebody really cares and edits the AI text, the AI output, they're probably going to be able to fool us. Okay? So that's what we're going to try on later on in this video try out, right? We're going to go ahead, we're going to do those little tricks that we explored last time. We're going to add a little space, see if that works. Okay, so that's the second one. And here I want to add Sam Altman himself talked about this a few days ago and in multiple other talks, I kind of listen to all of them, have to admit, he talked about the fact that a determined user will always be able to fool AI detection software. All right, so essentially this is a new world. That's his words, and it's completely true. This is a completely new world we just entered. So none of this software is going to save university professors or anyone else from AI generated text, okay? This is nothing but a recommendation, and it's just simply going to catch the lazy copy pasters, okay? So you have to adjust to the new world. If somebody's determined enough, they're always going to be able to get around this. Okay? So keep that in mind. And my last bullet point that I want you to know before we start testing this is. Well, it says right here, it says on their website, it makes it very clear, this works for English only. So sorry, to all the other people that speak Deutsch or any other language, this is for English only, okay? And it's again, kind of a research. It's a preview, right? It's a preview. They don't say this is a definitive tool. So with all that being said, you can look into the details. The link is in the description below. I want to just get right into this. I want to get into this and start testing it. But very last thing we can do before we test it with our own outputs, I prepare that email. I prepared some tricks that I'm going to throw at it, some curved balls. And then we're going to try and engineer and get around the detection and make it like, you know, non AI, although it has been generated by AI. So after we do all that, we'll take some questions and some suggestions from you guys in the comment section here, so feel free to leave them below in the comments in the chat, as this is live, which I kind of enjoy, to be honest, because now we can interact. So. So once all that is done, we'll be finished with the video. So with all that being said, let's try the little test interface they gave us here. After I take a little sip. Man speaking is exhausting. All right, so the first thing we can do here is just human written text. All right, so if I say submit, let's zoom in a little bit so you can see this better. That looks good to me. That looks good to me. Okay, so as you can see, I. If you hit that, okay, it is going to give you a recommendation. It's going to say, very unlikely AI Generated, okay? That means this was probably not generated by AI. And down here below, you can see what this exactly means in terms of probabilities and all. Okay, so very unlikely is actually the lowest tier. So this is very good. Okay, Very unlikely to be AI generated means human wrote this. The thing is, there's false positives and all that. So it's like, look, it's not perfect. Okay, I think we clarified that by now. False negatives. My bad. Anyway, so if we go to AI generated, okay? And say submit, it says possibly AI generated. Now let's review the table here below. Okay, Possibly AI generated. Hmm. That is the second worst category in terms. Okay, so, so look, it's even in their example when you feed it. This is. They want to communicate something yet again, Right? Even if you feed it clearly AI generated text. I mean, that's their damn preset. Okay? It doesn't even go fully clear on you. It tells you it's possibly generated by AI. Right, which is the second category from the bottom. All right, and, and with that being said, the last example here is an example of an incorrect classification of human written text as AI generated. So a human wrote this and it thinks it's written by AI just the way it is. Okay, so this is a Spanish example. Again, don't use this for languages outside of English. So this is likely generated by AI. Okay, so what we're going to do here is very simple. We're going to go head on over to ChatGPT and we're going to tell it the most basic of all prompts. Okay, Write me an essay. Let's do a 200 words essay. Right? So we have the required length about penguins, because penguins are awesome. Okay, so now they're going to do that. We're going to copy paste this and then we're going to see how good it is at detecting it. I mean, it doesn't get more basic than this. Right, Next up, we're going to go into styles, because styles have been a little sticky for the software. So let's see what we get here. All right, I'm going to. I'm going to get into the questions and everything afterwards because, yeah, we want to focus on a task at hand here, all right, and deliver you these, this information about, you know, penguins. So, okay, so it finished. Fantastic. I'm so glad it's live right now. That would be a little bit of a bummer. We have Playground as a, as a backup here. But let's just go with straight chatgpt here, okay? And let's feed it to the AI text classifier until it submits. So first of all, it's wonderful that it just works and it says possibly AI generated. Okay. All right. So this is the first one. So it kind of caught it, right? It kind of caught it. Let's be real. That's the same classification, if you remember the second one from the bottom. Yeah. So somebody, if somebody could clarify. Look, I'm not a researcher or a statistician. Am I correct if I presume that 0.9, between 0.9 and 0.98 means that 90%, there's a 90% probability or like that it's fake, that it's AI generated? Is that correct? Am I interpreting that correctly? If somebody could clarify, that would be fantastic. Either way, this is like if you get possibly AI generated, it's probably generated by AI. Okay. So if somebody sends your email and you. And it's kind of like, it's a little too good to be written by that person. You just copy, paste it in here and it's going to tell you possibly. Or like you can just screenshot. You can just screenshot this and send that to the guy. A nice little comeback right there. Either way, we test the first thing. All right, so now that we did this, I want to add a style. This is gonna be my first curved ball added. Alright. Okay, so we're gonna add in penguins in the style off. And now we're just gonna say, yeah, it means 90%. All right. Yeah, that makes sense. In the style of. Let's do something like, pretty usual here, right? We'll get into the esoteric edge cases soon here. Don't worry, don't worry. I love those. I love myself in the style of Borat. That's just perfect. So in the style of. I'm blanking here. Let's just do, I don't know, John Oliver. Okay. Okay, ladies and gentlemen, let's talk about penguins. So this is like, it simulates the style of a human being, which, you know, in theory, just looking at this, that should be a pretty good way to fool it. But then again, I don't know how this has been trained, how it has been prepared. So let's see what we get here. Okay, so this is going to be exciting. What do you guys think? What do you guys think? Predictions Maybe we could do. Wait, can I, can I get a little, I can, I can do a little poll. Okay, I'll do a little poll on the next one. That's actually pretty cool. We can do like little predictions here. Live for now. I'm just gonna copy this. Okay. I think this one it should catch, right? Like, it's just a basic style, but then again it like simulates a human being. So we don't know. Let's see. Submit. Unclear. Okay. Unclear. If it is generated by AI. If we refer to the categories, that's kind of a middle of the road type of answer. So it wouldn't know. So that's interesting. So this has clearly been generated by AI, but I told it to mimic the style of a human that is, you know, well documented out there in terms of their style. So it's not clear now I want to move on to the next one. And as I run this, write me and 200 word essay about penguins, of course. And no, and I want to do this. In a style that is not typical to AI. Okay. I came up with this prompt and I had it in a video and actually went viral on TikTok, funnily enough. So as I submit this, we're going to start a little poll here. Okay. We're going to have a little fun with this. Is it going to catch this one? So vote yes. Yes and no. Okay. All right, the questionnaire is out. You can go to the chat and you can vote. Is it going to catch this one? If you think it's going to catch this. So my definition would be if it gives it one of these two, it caught it or. All right, so if it says likely or possibly AI generated, and if it gives it one of these three, it did not catch it. Right. So what do you guys think? What do you guys think? It's not going to catch it, right? I mean, hey, GPT is pretty damn smart here and we're literally asking it here to, you know, generate it in a style that is not typical to AI. So you have a few more seconds here to vote and we're just gonna. Just gonna go ahead and test this out. So, you know, from, from all my previous tests, this fooled all the GPT2 detectors. This kind of like destroyed them in a sense. So let's see how this does. Delete the last one. Paste. Okay, let's see what the results are here. 71% think no. Okay. This is actually pretty fun. I love this. Okay. 70%, 68% think. Okay. So two thirds of the people think it's not going to catch it. So let's test it out. Let's do this. Submit. Unclear. Oh, boy. Oh, boy. That means I. I would have said no too. I would have said no too. I set up the rules. Most of us are wrong. It kind of. It didn't clearly catch it. Wait, we're right. My bad. Being live is hard. Okay. Yeah, we're right. So 2/3 of the people said no. Me included. So if it. If it's in one of these category, it's a no. Right. It didn't catch it, but it's still got the unclear. So, you know, all right, most people here got it right. But got to give some props to this detector. It's not bad. Even the. In the style of a human being and in the style that is not typical to AI, it's. It gave it the unclear rating. Right, let's see. So, okay, so let's move to another one here. Okay, here's an email that I wrote. This is going to be our next text test. Here's an email that I wrote. Right. This is a part of the welcome sequence if you subscribe to my newsletter, which, by the way, is a fantastic idea. And it's essentially like my life story where I just talk about how I learned how we moved, like every three years and I just had to learn a bunch of new skills. And AI is another one of those. And that's how I got here. So we're going to post this. Okay, we're going to post this to the detector. This is a completely fully handwritten email. Okay. And okay, so this is the handwritten email and we're going to run it through it. So I'm curious. I don't know. We'll do a poll on the next one. All right, so let's do submit. Let's see what it does here. So this is like, purely, like human written. Let's see what it says here. Okay, it doesn't. Maybe let's just refresh. It's not working right now. It's gonna be down a lot. Just like ChatGPT. It might be. Oh, please enter at least 1,000 characters. All right, all right, no problem. Oh, no problem. Except I kind of, you know, close the email. So give me, like, I have it up already. Okay, this is good. Give me like a few more seconds and I'll just, you know, open this up. So I'm gonna open up a longer chunk. Okay, perfect. So this should be enough. Let's see if it detects my handwritten email. If it detects. I mean, if it classifies it correctly, unlikely AI generated. Okay, so let's refer to this. Honestly, I kind of still. I'm not used to it yet. So I always need this. This charts down here. Unlikely. Okay, so it's. Okay. Fair enough. Fair enough. To be completely fair. Actually, now, thinking about it, I think there's maybe there might be like, one or two sentences where I used AI to inspire some of this. Yes. Because I, like, wrote up the. And this is how I use it. This is my quick tip here. I just wrote out the whole story by myself, and then I told AI to rephrase it and, like, various styles, and then I read through those and used it as inspiration. I didn't even copy just to go back and rewrite some of the sentences until I liked it more. Right. So. And I believe I rewrote, like, maybe two or three sentences, now that I think about it. Yeah. But I didn't, like, copy paste anything, so actually, that is pretty damn accurate in that case. I mean. Okay, maybe it's not 10 to 15. About 10, 15% of human written and 10% of this label. Okay. So unlikely. So that's really good. I think if you rank in one of these two, it's just like, okay, human wrote this. We're safe here. So let's move on to the next test here. Let's move on to the next test here. This has been an interesting one. Okay, so that's really good. I'm really glad because I put, like, some thoughts and this was, like, legitimately human written, you know, so I'm glad I didn't classify this as AI written. So let's go with a little bit more of an esoteric style here. Okay. And I want to do a poll here. So we're gonna do in the style of Borat saga, df. All right. I think that man needs no introduction. Is it going to catch in the style of Borat? Let's see what you guys think. I like. I feel like this is. This is such a hard one. So vote yes if you think that once I paste it, it's going to classify it as possibly AI generated or likely AI generated. And vote no if you think it's going to be one of these three. Okay, poll is out. Let's do this. Let's generate some Borat Penguin essay text here. It's one of my favorite prompts. Honestly, it's limited because it just doesn't have that much source material from Borat, but nevertheless, it's fantastic. Let's see. In a stylo. Okay, whatever. That's the beautiful thing. Like, you don't have to be perfect in your inputs. It just does it. Okay, let's see. What do you guys think here? I'm Really? I really don't know. I like it, honestly. This is. This is kind of a coin flip, isn't it? Okay, let's see. Okay. 61 of the people think it's gonna catch it. Really? Like, I don't know. The more I think about it, like, Borat really going to be like, penguin is so very nice, you know? And then, like, how. How is it going to know AI wrote that? I mean, it kind of makes sense. Like, these styles. It just feels like these styles are so hard to detect. Right. Especially if you take, like, characters like this. Okay, I'm curious. I'm really curious for this one. I don't know about you guys, but this is kind of the one that I've been waiting to explore. Also, none of this has been, like, prepared or scripted, so all we're doing is, like, just testing things that I'm curious about here. And I haven't even tested the classifier yet. I think I ran it, like, once on this, on the free examples. That's it. So let's see. It's not really cooperating here. Should we just reload? I think that's going to be the best. Okay, thanks for all the comments. Appreciate it. Guys, I'm reading in. Let's just try again here. And in case this doesn't want to work, let's just do, you know, let's just try Playground maybe, because this seems overloaded right now. Okay. And I, Borat, would like to share some facts about them with you. Okay. So this is not really in the style of Borat. If I know something, then it's. It's the style of Borat. This is not the style of Borat. No, no, no, no, no. This is this. I don't accept this output. Come on, chatgpt. Give me something. Defining Slavic. No, that's not. I meant it's a video on my channel where I talked about it. Wait a second, I'll show you. I'll show you. We talked about it here. So I think I messed up a little on the titling and everything here, but it's this one. Chat GPT writing detectors are useless. I tested the GPT2 writing detector in various ways, and it just, like it completely fell apart. This one I'll give you. That's a too long TLDR on it. And then the other one where we tested them is. This one. How to get around ChatGPT. So this was my first attempt, and here I went way deeper, and it just fell apart once I went deeper. Okay, we're not getting this Borat output here. So I thought about this happening. So we'll do one more try and then we'll just move on to something that playground can generate. You know, style of Borat is just a little too advanced for ChatGPT here. Makes sense. Makes sense. Look, in the style of the Borat. No, actually, not at all. We can try that. But yeah, it doesn't change anything. It actually doesn't. It doesn't care about minor grammatical mistakes. It just doesn't. Do this tool. But yeah, it looks like it's overloaded here. It looks like it's overloaded here. So let's just try once more. But otherwise we'll just move on and we'll. We'll go over to my next request here and we'll just take this. We'll just take basic AI generated text. Okay. So this is going to be my first big test. Sorry, it was a 50 50. It was a straight 50 50. So we'll get back to this one in the end. That's a good one to end things with. But we'll just take our basic penguin example here. Close browser and reopen. Yeah, but that would like, close all my live streaming stuff here too. I can. I suppose I can do this. Okay, we'll take a basic penguin text here. Okay. And then will be like, now rephrase the following text to be more conversational. Right? Maybe conversational text. It shouldn't catch us easily. And then it should do that. This at the bottom. So we're going to take the rephrase text from GPT3 here and we're going to test that. Yeah, ChatGPT seems to be stuck right now. So it's fine. It's fine. Wanted to check both. Both. Even the most unlikely of creatures can overcome. Okay. Penguins are undoubtedly one of the cutest creatures around. So this is good. This is really good. It might be a little short. Right? Let's see. Okay. Submit at least thousand characters. Okay. More conversational. And extend it to be twice as long. All right, so once we do this, this is going to be an interesting one because the biggest way of getting around these detectors is, first of all, rephrasing. And secondly, using external app apps like Quill Bot. And that's the next one we're going to do here. Okay. We're going to use the paraphrasing tool Quillbot to rephrase the penguin text. That's going to be our next test here. Okay. So it's still a little short, right? Still a little short. Let's do let's up the tokens here. Let's try one more time. So these are the interesting tests, right? This was like a warm up up until now, but now we're getting into the meat of it. Rephrasing and Quillbot are the two. It's a little longer, but yeah, honestly, dealing with length is one of the biggest. Yeah, it's just the biggest limitations here. It's one of the biggest limitations. Yeah. Chat GPT is down. Let's see. Let's try again. Ah, there we go. There we go. Okay. Borat style. It's back on. It's back on. I'm so glad this works. Okay, we're gonna see if the penguin essay in the style of Borat. Very nice. I'm sorry, I'm sorry. I apologize. I shouldn't be doing that. I shouldn't be doing it. It's just he's. I don't know. He's too great. He's an all time favorite of mine. He was the very first DVD I ever bought when I was like 12 or something. Okay, Nana, perfect. Okay, well, let's take this and let's paste this into the classifier and let's see if it catch. Oh, no, no, no, no, no, no, no. You can always tell it. Write more. Right. By the way, this is a little pro tip here. Write more and it's going to extend a little bit. I feel like we need like maybe two or three more sentences for this to work. At least 1,000 characters. Yeah, 200 characters more. You know what? You know what? We're gonna cheat a little here. We're gonna cheat and we're gonna just. We're just gonna take like two sentences here and we're gonna add it to the Bora Text. Gonna add it to the Bora. Okay, let's do this. This should be fine, right? It's all AI generated, so it should catch it. Let's see. Unlikely. AI generated. Wow. Wow. A good old. No, no, no, no. Folded. So what do we do from here? Well, clearly we're going to keep playing here. We're going to keep playing here and we'll take our very basic Penguin essay and we're going to tell ChatGPT, rephrase this. And maybe, maybe let's do in our conversation. Okay, perfect. In a conversational style. So nice. It's working. So we're going to take the basic write me an essay prompt rephrased into a conversational style. It should. It should catch this. If it doesn't catch this, I would lean towards maybe don't even use the tool, honestly. Because if you. If you write anything and then you just tell it. If the workaround is like, just tell it. Rephrase once. And I know this is. Look, this is no scientific study, right? Come on. We're just a guy with a neon flame. Flame in the background, testing these things, how they work in practice. But honestly, if it doesn't catch, the rephrase prompt feels kind of useless. I hope it does, and I think it will. All right. Okay. So Borat, Paul poll has ended 50 50. And on this one. Let's do it. Let's do it. Let's do it. I'm really curious. This is kind of. This is maybe even the most important one. So are you ready for this? I'm actually kind of. I'm kind of nervous because if this. If it doesn't catch this one, it's. It's not looking good. Possibly AI generated. That's good. That's good. Okay, so it caught it. So it seems like, okay, that's really good. So it doesn't work. So you can't just be like, okay, let's. Let's do one more thing. Now rephrase this in a style that is not typical to AI. Okay, so last time it caught this, but now we're doing again. We did an essay, then we rephrased it, and then we rephrased it in a style that is not typical to AI. So this, like with the other detectors, I was able to break all of them with this prompt. Okay. Okay. And I, Borat, would like to share some facts about them with you. Okay, so this is not really in the style of Borat. If I know something, then it's. It's the style of Borat. This is not the style of Borat. No, no, no, no, no. This is this. I don't accept this output. Come on, chatgpt, Give me something Borat. Like here, Defining Slavic. No, that's not. I meant it's a video on my channel where I talked about it. Wait a second. I'll show you. I'll show you. We talked about it here. So I think I messed up a little on the titling and everything here, but it's this one. Chat GPT writing detectors are useless. I tested the GPT2 writing detector in various ways, and it just, like it completely fell apart in this one. I'll give you. That's a too long TL doctor on it. And then the other one where we tested them is. This one. How to get around ChatGPT. So this was my first attempt, and here I went way deeper and it just fell apart once I went deeper. Okay, we're not getting this Borat output here. So I. I thought about this happening. So we'll do one more try and then we'll just move on to something that playground can generate. You know, style of Borat is just a little too advanced for ChatGPT here. Makes sense. Makes sense. Look, in the style of the Borat. No, actually, not at all. We can try that. But yeah, it doesn't change anything. It actually doesn't. It doesn't care about minor grammatical mistakes. It just doesn't. Do this too. But, yeah, it looks like it's overloaded here. It looks like it's overloaded here. So let's just try once more. But otherwise we'll just move on and we'll. We'll go over to my next request here and we'll just take this. We'll just take basic AI generated text. Okay. So this is going to be my first big test. Sorry, it was a 50 50. It was a straight 50 50. So we'll get back to this one in the end. That's a good one to end things with. But we'll just take our basic penguin example here. Close browser and reopen. Yeah, but that would, like, close all my live streaming stuff here, too. I can. I suppose I can do this. Okay, we'll take a basic penguin text here. Okay. And then we'll be like, now rephrase the following text to be more conversational. Right. Maybe conversational text. It shouldn't catch us easily. And then it should do that. This at the bottom. So we're going to take the rephrase text from GPT3 here, and we're gonna test that. Yeah, ChatGPT seems to be stuck right now. So it's fine. It's fine. Wanted to check both. Both. Even the most unlikely of creatures can overcome. Okay. Penguins are undoubtedly one of the cutest creatures around. So this is good. This is really good. It's gonna. It might be a little short. Right? Let's see. Okay. Submit at least thousand characters. Okay. More conversational. And extend it to be twice as long. All right, so once we do this, this is going to be an interesting one because the biggest way of getting around these detectors is, first of all, rephrasing. And secondly, using egg creatures can overcome. Okay. Penguins are undoubtedly one of the cutest creatures around. So this is good. This is really good. External app apps like Quillbot. And that's the next one we're gonna do here? Okay, we're gonna use the paraphrasing tool Quillbot to rephrase the penguin text. That's gonna be our next test here. Okay. So it's still a little short, right? Still a little short. Let's do. Let's up the tokens here. Let's try one more time. So these are the interesting tests, right? This was like a warm up up until now, but now we're getting into the meat of it. Rephrasing and Quillbot are the tool. It's a little longer, but yeah, honestly, dealing with length is one of the biggest. Yeah. It's just the biggest limitations here. It's one of the biggest limitations. Yeah. ChatGPT is down. Let's see. Let's try again. Ah, there we go. There we go. Okay. Borat style. It's back on. It's back on. I'm so glad this works. Okay, we're gonna see if the Penguin essay in the style of Borat. Very nice. I'm sorry, I'm sorry. I apologize. I shouldn't be doing that. I shouldn't be doing it. It's just easy. I don't know. He's too great. He's an all time favorite of mine. He was the very first DVD I ever bought when I was like 12 or something. Okay. No, no, no, no. Perfect. Okay, well, let's take this and let's paste this into the classifier and let's see if it catch. Oh, no, no, no, no, no, no, no. You can always tell it. Write more. Right. By the way, this is a little pro tip here. Write more and it's gonna extend a little bit. I feel like we need like maybe two or three more sentences for this to work. At least 1,000 characters. Yeah, 200 characters more. You know what? You know what? We're gonna cheat a little here. We're gonna cheat and we're gonna just. We're just gonna take like two sentences here and we're gonna add it to the Bora text. Gonna add it to the Borat. Okay, let's do this. This should be fine, right? It's all AI generated, so it should catch it. Let's see. Unlikely AI generated. Wow. Wow. A good old na na na na na. Fooled it. So what do we do from here? Well, clearly we're going to keep playing here. We're going to keep playing here. And we'll take our very basic Penguin essay. It might be a little short. Right? Let's see. Okay. Submit at least thousand characters. Okay. More conversational and extend it to be twice as long. As long. All right, so once we do this, this is going to be an interesting one because the biggest way of getting around these detectors is first of all, rephrasing. And secondly, using external app apps like Quill Bot. And that's the next one we're going to do here. Okay. We're going to use the paraphrasing tool Quillbot to rephrase the penguin text. That's going to be our next test here. Okay. So it's still a little short, right? Still a little short. And we're gonna tell chatgpt rephrase this and maybe, maybe let's do in our conversation. Okay, perfect up the tokens here. Try one more time. So these are the interesting tests, right? This was like a warm up up until now, but now we're getting into the meat of it. Rephrase in a conversational style. So nice, it's working. So we're going to take the basic write me an essay prompt rephrased into a conversational style. It should, it should catch this. If it doesn't catch this, I would lean towards, maybe don't even use the tool, honestly. Because if you, if you write anything and then you just tell it. If the workaround is like, just tell it, rephrase once. And I know this is, look, this is no scientific study and Quill Bot are the two. It's a little longer, but yeah, honestly, right. Come on. We're just a guy with a neon flame, the flame in the background, testing these things, how they work in practice. But honestly, if it doesn't catch the rephrase prompt feels kind of useless. I hope it does, and I think it will. All right. Okay. So Borat Paul poll has ended 50 50. And on this one. Let's do it. Let's do it. Let's do it. I'm really curious. This is kind of, this is maybe even the most important one. So are you ready for this? I'm actually kind of, I'm kind of nervous because if this, if it doesn't catch this one, it's, it's not looking good. Possibly AI generated. That's good, that's good. Okay, so it caught it. So, so it seems like, okay, that's really good. So it doesn't work. So you can't just be like, okay, let's, let's do one more thing. Now rephrase this in a style that is not typical to AI. Okay, so last time it caught this, but now we're doing. Dealing with length is one of the Biggest. Yeah, it's just the biggest limitations here. It's one of the biggest limitations. Yeah. ChatGPT is down. Let's see. Let's try again. Ah, there we go again. We did an essay, then we rephrased it. And then we rephrased it in a style that is not typical to AI. So this, like with the other detectors, I was able to break all of them with this prompt. Okay, let's see if it works this time. So honestly, up until now, this looks like by far the best detector I've tried yet. Quite reliable, actually, just from this little test. Submit unclear if it is AI generated. Okay. So we managed to throw it off here, but it's still in the middle. Right, There we go. Okay. Borat style. It's back on. It's back on. I'm so glad this works. Okay, we're gonna see if the Penguin essay in the style of Borat. Very nice. I'm sorry, I'm sorry, I apologize. I shouldn't be doing that. I shouldn't be doing it. It's just easiest. It's not. It's not saying very unlikely to be or unlikely. So it's like, this is good text. Okay, so. So we have a suggestion here. What adding. When it comes to writing, two factors are crucial, Perplexity and burstiness. Perplexity measures the complexity. Okay, so we should add. Okay, let's try this. Let's try this suggestion. So let's say now make it more perplexed. Right? Is that what you meant? This is. This is a good writing style. So is this what you meant at Envy? Making it a little more perplexed. So next up, we're gonna do. We're gonna do Quill Bot, right? That's the big one. That's the one that has been trending on TikTok and all that. Hmm. Curious to see. I don't know. He's too great. He's an all time favorite of mine. He was the very first DVD I ever bought. When, if, if it catches this again, for everybody tuning in, we're covering the brand new tool by OpenAI, official plagiarism detection tool. Kind of, you know, like, like AI detection tool. I think that's more accurate. The AI text classifier. You can paste text in here and then it tells you if it's AI detected or not or what it thinks. So we made it a little more perplexed. Let's see if it catches that. Submit Unclear. Okay. Still unclear. Okay, so let's switch gears here a little bit, ladies and gentlemen. Okay, let's take the very basic penguin text, right? The penguin essay, the beloved example on this channel. And let's actually run this for Quillbot. Okay, so this is the tool that all those Tiktokers, you know, have been, like, recommending. Yo, you have to check out this tool that rephrases like AI text. Now you can make like $100 a day, you know, pasting blog post. That's what all those TikToks go like. Right? I should get more. More into TikTok. Seems like a good place to learn. It is, though there's some corners of it that are good to learn. Anyway, I digress. Let's run this. Okay, so it's not possible. So we're gonna do it in two parts. Okay, we're gonna do it in two parts. Let's do just simple, standard paraphrase. Okay, there it is. There it is. So I'm just going to post second part here. We're going to get around paying for this little test demo. Otherwise, it's a. It's a worthy app. But right here, that's the first part. And then if we take it from here, we're going to able be able to do this. And you know what? We'll just. We'll just delete a few words. Oh, God. We can't. Okay, we're gonna have to do this in free runs. No problemo. So it's paraphrasing it. And now magical donut. What would happen if you copy paste it directly from Wikipedia? That's a fantastic question right there. I agree. That's a fantastic question. We're gonna do that. And this is why I'm excited. I was like 12 or something. Okay. Na, na, na, na, na. Perfect. Okay, well, let's take this and let's paste this into the classifier and let's see if it catch any. Oh, no, no, no, no, no, no, no. You can always tell it. Write more. Right. By the way, this is a little pro tip here. Write more and it's going to extend a little bit. I feel like we need like maybe two or three about this live format, because that's a great idea that I would have left out here. Okay, so this is the second part, and this is the last part. Generation be more sentences for this to work. At least 1,000 characters. Yeah, 200 characters more. Okay, I think. I think that makes sense. Let's just say paraphrase. Okay, so this full thing has been paraphrased, right? Let's see, let's see. Submit, right? Yes. Looks good. It's complete submit. Unclear if it is AI. This thing is really good. This thing is really good. Think about that. We. We ran it for like 10 tests, right? And I feel like not a single time did it like completely misinform us. Right? On all the stuff that was clearly gener. You know what? You know what? We're good by AI and simply copy paste it. It was just like. Yeah, it's. It's possibly AI generated on the one thing where it was. Maybe we should have more like personal writing examples. But on the one email that was like handwritten by me, it told us that it's unlikely, so it got that too. And on all the stuff where we tried to be tricky, the rephrasing the stuff. Cheat a little here we're going to cheat and we're going to just. We just going to take like two sentences here and we're going to add it to the Borat text. Going to add it to the Bora. Okay, let's do this. This should be fine, right? It's all AI generated, so it should catch it. Let's see. Is on all that it told us, it's unclear if it is written by AI. Interesting. Interesting to say the least. Let's do another one here. So let's do Wikipedia. Unlikely AI generated. I love that idea. Let's do Wikipedia. Thanks for the suggestion. Wow. Thanks for the suggestion. Magical donut. I mean, wow. A good old. No, no, no, no, no. Fold it. So what do we do from here? Well, clearly we're gonna keep Fantastic name right there. Man has good questions. So let's just take AGI here and. Oh, wait, this is German Wikipedia. Wait, it doesn't work in German. So we should. We shall switch this to English. Perfect. AGI, let's just take the first paragraph from Wikipedia if it. But this is not AI generated. Right? So this is. Yeah, I don't know. I don't know what. I don't even know what we should get here. Wait a minute. I'm getting a little lost here. There we go. So, okay, delete Wikipedia article with the references and all submit. Oh, at least thousand. Okay, so we need a little more. We're just gonna do this. Are you guys enjoying this? Is this a good test? I. I felt like this would be valuable content right off the bat. Maybe you have assignment due tomorrow and this is exactly what the doctor ordered there. Don't use AI AI to cheat, boys and girls. It's a fantastic source of inspiration. But that's all it All. All. It's Good for now. It's good for now. Unclear. Okay, so, so tell. It says it's unclear. There you go. So it's again the middle of the road thing. So it doesn't clearly tell you it's AI, which this is not AI generated, right? Wikipedia is not AI generated. But honestly, if it's this clearly copied from. It would be cool if this thing had a feature where it tells you, oh, no, no, no, no. This is like clearly copied from this website and then it gives you a link, right? That would be a nice upgrade to this tool. But I don't know, for now, it tells you like it's unclear. So look, the only example of human written text I fed it, it kind of clearly. It clearly stated that, you know, it's human written. So that's interesting. Do longer prompts not make it mess up? Messed up playing here. We're going to keep playing here and we'll take our very basic Penguin essay and we're going to tell ChatGPT, rephrase this and maybe, maybe let's do in a conversation. Okay, perfect. In a conversational style. So nice. It's working. So we're going to take the basic write me an essay prompt rephrased into a conversational style. It should catch this. If it doesn't catch this, I would lean towards. Maybe don't even use the tool, honestly. Because if you. If you write anything and then you just tell it if the workaround is. Well, I think I don't want to go too deep here. I just wanted to make this a first test, but I think from what we saw, honestly, let's summarize. Okay, so when we fed it the text straight up, okay, write me an essay. It stated, yes, this is AI. This is probably AI. If we fed it a style like just tell it. Rephrase once. And I know this is. Look, this is no scientific study, right? Come on. We're just a guy with a neon flame in the background testing these things, how they work in practice. But honestly, if it doesn't catch the rephrase prompt, Phil, it clearly read the fact that it was kind of useless. I hope it does, and I think it will. All right. Okay, so Borat, Paul poll has ended 50 50. Simply, a style applies to AI writing and it identified it too. And on this one. Let's do it. Let's do it. Let's do it. I'm really curious. This is kind of the next one we ran is my little tricky prompt. Write it in a style that is enough. This is maybe even the most important one. So are you ready for this? I'm actually not typical to AI. It caught that too. Then I fed it a personal email and I'm kind of nervous because if this, if it doesn't catch this one, it's. It's not looking good. It told me it's probably human written. Then we went possibly AI generated. That's good. That's good. Okay, so it caught it. So it seems like, okay, that's really good. So it doesn't work. So you can't just be like, okay, let's do one more thing. And rephrased it with the rephrase keyword here. And it told us it's not sure, middle of the road. But it didn't tell us it's human written. And in the end when we went with like a Wikipedia copy paste, it told us the same thing. Now rephrase this in a style that is not typical to AI. Okay, so last time it caught this, but now we're doing. And the final boss, Quillbot. Again, we did an essay, then we rephrased it, and then we rephrased it in a hidden. It didn't. It didn't fully catch it, but it didn't consider it to be human written. Whereas all the other detectors up until now here style that is not typical to AI. So this, like with the other detectors, I was. They always felt, they always fell apart. They were like, yeah, this is. This is human written. But he's able to break all of them with this prompt. Okay, let's see if it works this time. So honestly, up until now, this looks like by far the best detector I've tried yet. Quite reliable actually. The actual OpenAI detector. I gotta say, Hats off, ladies and gentlemen. Hats off. This thing is surprisingly solid. Surprisingly solid. So unless there just from this little test submit unclear if it is AI generated. Okay, so we managed to throw it off here, but it's still in the middle. Right? There are any further requests or questions, I think we'll round it out here. It's not. It's not saying very unlikely to be or unlikely. So it's like, this is good text. Okay, so. So we have a suggestion here. Thank you so much for joining me. Thank you so much for joining me in this live stream. I think I'm gonna be doing more of these when news worth what adding. When it comes to writing, two factors are crucial perplexity and burstiness. This stuff comes in and as you know, the AI space is a space that moves faster than complexity measures the complexity okay, so we should add. Okay, let's try this. Let's try this suggestion. So let's say now make it more perplexed. Well, probably any other space within technology right now. There should be newsworthy stories coming soon. Thanks for joining me and I'll see you around. Right? Is that what you meant? Good to have. This is. This is a good writing style. So is this what what you meant at mv, making it a little more perplexed? So next up, we're gonna do. We're gonna do Quill Bot, right? That's the big one. That's the one that has been here. El Couch Slavic. Thanks for your suggestions. I don't want to go too deep and test code. I'm trending on tick tock and all that. H. Curious to see. Feel free to play with it yourself. The text, the link is in the description below. Have a great day, everybody. Have a great day. If. If it catches this. Again, for everybody tuning in, we're covering the brand new tool by Open AI Official Plagiarism detection Tool. Kind of, you know, like. Like a. Eye detection tool. I think that's more accurate. The AI text classifier. You can paste text in here, and then it tells you if it's AI detected or not or what it thinks. So we made it a little more perplexed. Let's see if it catches that. Submit. Unclear. Okay. Still unclear. Okay, so let's switch gears here a little bit, ladies and gentlemen. Okay, let's take the very basic penguin text, right? The penguin essay, the beloved example on this channel. And let's actually run this for Quillbot. Okay, so this is the tool.