I tried to hack the GPT-2 Output Detector with various tricks to accept my AI-written text as real.
———————————————————————————
Free AI Newsletter:
E-Book with 400+ ChatGPT Use Cases:
Discord:
Twitter:
———————————————————————————
VIDEO EQUIPMENT I USE:
VIDEO EQUIPMENT I USE:
Main Cam:
B-Cam:
Lens:
Mic:
Mic Arm:
Audio Mixer:
Camera Switcher:
Light:
Main Computer:
———————————————————————————
#chatgpt #ai #aidetector
Apparently that little change made this Entire text legit Identifying AI written text right now is No easy task but in a lot of cases one Has to tell if this came from a human or AI model for example teachers evaluating Students and universities and in one of My last videos we looked at the gpt2 Detector that can apparently tell AI Written text from Human written text but The comment section flooded with people That wanted me to test this further and Rightfully so with this Channel all I do Is overuse some of these AI tools and Then share my findings with you guys Here so with the detector it seems like We need to go a little deeper and that's Exactly what we'll do today we'll run it For a total of 15 texts some of them Being legit some of them being AI Generated and then we'll evaluate its Core and how accurate it has been by the End of this video I hope to find out how Relevant the results from this detector Actually are but whatever the outcome Might be the gpt2 output detector is the Best freely available tool to identify AI generated text that consumers have Right now so without further Ado let's Start with the most basic thing Generating text with AI and copy pasting It into the detector alright so now it's Time to copy paste this right into the Detector let's see what we get okay
Super clean detection 99.91 fake and in This case the definition of fake is a Little loose in terms of coming from Ai And you know what might as well have Some fun while we're at it oh whoa whoa Whoa it says that's real so apparently If you choose the voice of Borat it Thinks it's real alright so it doesn't Seem to handle Styles very well so let's Try the same thing without the style Okay it caught that okay so it seems Like every time you prompt it straight Up it catches it but just to make sure Let's do the same thing but now add in a Conversational style which is not as Exotic as asking for borats although I Do prefer that one okay copy paste it Says it's real look so the two ones Where we just asked straight up it Detected and every time we said in this Dialogue it didn't even catch it so Already that's not looking good for the Detector but let's keep moving here okay So here let's just take one of the Emails that I wrote today and copy paste It and see if it catches it all right if This says fake there's clearly something Wrong and there you go it says 95 fake This was written on my phone on the go So I can 100 guarantee you there was no AI involved so a big fan L for detector Here too but let's do one more email Alright this should do Copy paste and let's see what it says oh
And this one it thinks it's 85 real Again no AI evolved here but it still Thinks it's partially fake all right and What about this what if I say rewrite my Email in a more professional and concise Manner then I post the email so it even Include square brackets so this should Be pretty easily detectable right let's See okay just got two percent more fake Really okay let's do one more rewrite That and more of a conversational style So last time I didn't catch the Conversational style at all all right Copy paste and that just made it more Real and let's do one of my favorites Again now in the style of Borat me send You big hug I might just start using That myself honestly let's see if he Picks up on digital board here no he Doesn't at all okay so Borat seems to Pass this test every time for the next Test I want to do something completely Different I'm gonna copy all the text From my ebook here and I can already Tell you the only thing that has been AI Generated here is the copyright policy And the disclaimer and so from this 112 Pages only one is purely generated by AI Otherwise it just helped me in my Research so let's just paste 116 pages In here and see what we get okay so Apparently it's too long so let's just Delete half of it and there you go I had To trim two-fourths of the book in order
For this to pick it up but even with That copy pasted disclaimer it gave me 99.98 but then interestingly showed me That it only worked on the first fourth Of the texts so just be aware if you Wanted to copy paste the book into this Do it in multiple runs so if you find Videos like this valuable leave a like Below and sign up to my Weekly Newsletter it's completely free and Covers all things that are happening in The AI space right now okay next up We'll look at some of these tricks that I introduced the last time one of them Would be simply taking a text that the Detector identified as fake simply Saying now rewrite this not in the style Of AI copy paste okay so we went from 99 Fake to 40 fake not bad and here's Another thing I caught in the comments So apparently if you add a space at a Random spot it just stops working and There you go that actually worked look By adding one space it identifies this As completely unique so you can just add A space at any spot and get around this Detector and here I'll share something With you that is kind of on unethical But then on the other hand also teachers Watch this so watch out for this little Trick if you insert the space and you go Into Microsoft Word you can then Actually reformat the two spaces to be a Smaller font size to make it look like
One space when you print it but then it Will get around this detector right so As you can tell from this video this Thing is falling apart fast let's just Test the space trick one more time okay That's the first text we generated that Is completely AI generated and clearly Fake and let me just add a space at a Random spot and see what happens well Apparently that little change made this Entire text legit okay one more space See if we can clean this up and make This completely legit wow look at that Okay this is kind of addictive let's do This one more time and get this right oh No it got more fake by adding three Spaces interesting okay again the same Thing and what if I had a fourth space Okay it gets worse so it seems that two Spaces at random spots is The Sweet Spot Okay I guess we learned something and Honestly in the other video I even use Tools that rephrase the whole thing Which made it pretty legit in the eyes Of the detector and as this is the most Popular one from all the freely Available tools right now that leads me To the conclusion that this tool at this Point in time seems to not be a reliable Source whatsoever alright and if you Want to deepen your understanding of Chat GPT even further this video will Teach you some things that you might not Have seen before
