Today, I will be running AI-generated text
through various plagiarism detection tools. Including one that kind of specializes in
detecting ChatGPT outputs. Because let's be real - plagiarism is kind of the main
issue when it comes to AI-generated text. It's one of the first questions people usually
have: they have all these amazing outputs, but they're not sure how to use them. So, in a second,
we'll start by generating text and feeding it into a traditional plagiarism checker. Then we'll move
to an AI-specific one. But before we do that, you need to realize that, as the chatbots and the
AI gets better, so will the detection software. Copy-pasting outputs and presenting them as,
let's say, your bachelor's thesis is one of the worst ideas I've come across when using this
tech. There's just no way that, over time, these detection tools won't catch up with you. So,
while ChatGPT is unbelievably good and useful, especially if you know how to
communicate with it. Pretty much all of the time, it should only be used as a starting point or as
an assistant, not as a finished product. Just keep that in mind. But, with that being said, let's get
to generating some text here. So, today, we'll do two test runs: one with a super basic "Write
me an essay about the importance of avoiding plagiarism in universities". We stay on topic
here, okay? And the plagiarism detectors we'll be using here are just two results I found at the
top of Google; nothing special. So, one of them would be Duplicchecker.com and the second one is
Small SEO Tools.com/plagiarismchecker. Alright, and now let's copy-paste this text which talks
about the importance of not copy-pasting. Okay, so if I paste this text and check for plagiarism
in one and then in the second one, here are the results: the first one tells me it's 100% unique,
AKA it has no idea this was written by AI. And the second one also states it's 100%
unique, zero percent plagiarism. Okay, so these basic tools have no idea. I do have to
note that universities have more advanced models that going deeper and have wider databases. But I
can tell you from other people's experiences that these plagiarism checkers, even the university
ones, only catch about 10 to 25% of the text, which, in a lot of cases, is still
acceptable. So, as you can see here, traditional tools don't really catch onto ChatGPTs
output. But now we get to the AI plagiarism checker. This one is called the GPT2 outputs
detector demo. And it's hosted on a fantastic site called "huggingface" where people build their
own AI interfaces. You can explore some of these, and soon we're gonna talk about way more of them.
Because there are some real gems hidden in there; you just have to search and try around
a lot to find the good ones. Either way, GPT2 output detector is one of the good ones, and
you can find it under this URL. Oh, by the way, this thing is down more than it is up because so
many people are using it, I guess you'll just have to be patient. So, if you copy-paste the text
here, all I have to do is wait for a bit, and it's going to predict how much of this was written
by AI. And, right away, we have it: 99.97% fake. Okay, so this tool could expose this immediately,
even though this is based on the GPT2 model. And ChatGPT is the GPT3.5 model. So, even though
this is detecting the older model, it clearly catches this. But let's see if we can get around
this - can we fool this with some creative prompts engineering? Let's just try a basic: "Now write
it in a tone that is not typical to AI." Alright, so now let's copy-paste this new text that is
apparently written in a style that is not typical to AI. Oh my God, look at that - by simply telling
it: "Wight in a tone that is not typical to AI" I pretty much got around this version of the
detector. Alright, but here's another idea. And this line is straight from the 'how-to'
section of the second version of my ebook. So, what it does is that it teaches ChatGPT your
very own style, and then uses that to write. And, as a result, you should get a text that is
unique to your very own style, right? So, let's test this out, as per usual, with my
formulas. All you have to do is replace the part that is in square brackets. So, here I'm
just gonna paste a little bit of my own writing. Okay, now instead of "now rewrite this text",
I'll say "now rewrite your last response". Copy-paste. Ah, look at that - that didn't work
at all: 91% fake. But wait a minute - but we have one more trick up our sleeves. And that
is using some of the rephrasing technology that has been available since a minute.
So, let's ChatGPT, generate everything, and then you rephrase it in external
life. Let's see if that works, and this app is called 'Quillbot'. It's completely free
- you can just come in here and paste the text, and let's actually use the last text we had
here, and say 'paraphrase'. And Quillbot simply picks some alternative words here. Now
it's time for the final test - let's see if the AI detection tools can pick up on text that was
generated in my very own style and then rephrase. Whoa! 64.2% real. I mean, not bad - two-thirds
of this are real. And the thing is if you have access to these tools, which you do because
they're free, you could start altering this until you get a 90% real score. And then
it passes in most cases, right? But again, you just need to assume that these tools are
gonna get better: this is a GPT2 detector, and it catches almost everything from GPT3.5,
even once we rephrase this. So, while you might be able to get away with plagiarizing things
from a Chatbot today, there's just no way that's a good mid to long-term strategy, so keep that
in mind. Okay, but enough of all this talk that tells you how to not use ChatGPT - this video
will show you a bunch of creative ways on how to put ChatGPT to work in ways that you probably
wouldn't have imagined by yourself. See you there!