ChatGPT is so good at generating content in practically any area that many people have started to worry that they'll lose their writing jobs. I personally think this will happen.
In my opinion, people who choose to use AI to their advantage will become much better at their jobs. Those who don't might get left behind. But that's exactly what happened when the Internet became popular as well.
The Internet made it so easier to find information that you'd otherwise have had to spend days or weeks searching for by going through countless books at your local public library. Even then, you probably wouldn't get all the information you needed.
Anyway, I'm getting sidetracked here. People started massively using ChatGPT to write content. This has teachers worried that their students won't put the necessary work into research and write their essays. Clients also want to know whether the content writers they hired are actually doing work or using ChatGPT and sending then AI-generated text.
Naturally, there should be a way to check whether people are using AI tools like ChatGPT to write content.
Can You Check If Text Was Written By AI?
Yes, you can detect AI-written text. In this article, I'll talk about two tools that you can use to check if the content was written by artificial intelligence.
The AI Text Classifier by OpenAI
It didn't take long after OpenAI released ChatGPT to the public for people to become worried about whether they'll be able to make a distinction between AI-generated text and content written by humans.
Students, content writers, social media marketers, poets, and people in many other professions that include writing short or long-form text began to use ChatGPT for more than just inspiration and guidance. It can be problematic if you don't add even an ounce of your input to something a large language model generated.
It can also cause problems when someone hires you to do a job and you aren't being transparent about using AI technology to this extent. I also encourage students to learn as much as possible. Don't just copy and paste assignments. It's okay to use AI in a way that boosts the quality of your work. But don't let it make you lazy.
You probably get why there was a lot of panic regarding the detection of AI text. Thankfully, OpenAI stepped up and quickly released a free tool that enables you to figure out whether specific content was written by AI. This tool is known as the AI Text Classifier.
As you can see from the blog page on the OpenAI website about the AI Text Classifier, the company notes that the tool is still a work in progress. They write that the classifier is not fully reliable and can be especially wrong if you ask it to check a short text below a thousand characters.
OpenAI clearly mentions that the results you get from using the classifier can help you determine whether a document was generated with AI, but it shouldn't be used as the sole piece of evidence. It's entirely possible that the tool mislabels both AI-generated text and content written by humans.
When the classifier goes through your document, it will label it depending on how likely it is for it to be AI-generated. The labels include - very unlikely, unlikely, unclear if it is, possibly, and likely AI-generated.
Although I personally write all of the content in my articles (ChatGPT writes my meta descriptions though), I still wanted to check what the classifier thought about my content. I pasted my how to jailbreak ChatGPT article and used the AI text classifier.
As you can see from the image above, the AI text detection tool deemed that the content was unlikely AI-generated. In this case, that is correct. But I also wanted to use it to detect AI-written text by ChatGPT.
To do this, I asked ChatGPT to write a 1000-word article about the Mesozoic Era.
When I finished writing the article, I went back to the AI Text Classifier. I pasted the content and was eager to find out whether it would be able to detect it was written by AI.
The AI detection tool gave me a result saying that the text was possibly generated by artificial intelligence. In both cases, the classifier leaned toward the right answer. This is quite impressive in my book.
I have no doubt in my mind that OpenAI will take the time to properly train the classifier to become much better at detecting AI-generated text. In the meanwhile, there is another interesting AI text detection tool that you can use.
GPTZero - A Popular AI Text Detection Tool
GPTZero is an AI text detection tool that was developed by a Princeton student named Edward Tian. It was made available for free on January 2 and has since attracted a large number of users.
Tian made this tool primarily with educators and journalists in mind. Following the release of GPTZero, Tian announced that he was partnering with ed-tech organizations to give his tool access to a larger data set. This allows the tool to become more accurate and powerful.
There are two different properties that GPTZero measures to detect AI-written text. One is perplexity, while the other is burstiness. The former refers to the randomness of the text to the language model. In case the language model finds it random and unfamiliar, it suggests that a human wrote the content since it doesn't come from the data it was trained on.
The other measure refers to the complexity of the sentences in the text. ChatGPT will usually provide you with a consistent length of sentences. When humans write text, they'll almost always be chaotic. One sentence might be really short, while the next one may be much longer than needed.
To use GPTZero, simply navigate to the official website and you will see a box where you can paste the content you want to check. I'll do the same thing as I did in the section above. First, I will paste one of the articles I've written for this website.
Well, it got this one wrong. The whole text was written by me. But maybe I'm using ChatGPT so much for other stuff that I've started to write in a similar style. Just kidding, this just shows that tools like this can't be 100% reliable now. They need to be trained more.
I also wanted to give GPTZero the AI-generated text about the Mesozoic Era to see if it could detect it wasn't written by a human.
In this case, the results are absolutely correct. The text is written entirely by AI.
GPTZero started as a project Tian worked on during his winter break. It currently has more than a million users. It's quite impressive what he built in such a short time, so there's a good chance this AI detection tool will get much better over time.
When I compare the two AI detection tools I've written about in this article with the extremely limited set of examples, it seems that the AI Text Classifier by OpenAI is more reliable. However, I think you would probably benefit the most from using both of them for every piece of content you check.
How to Detect AI-Written Text?
There are easy ways to detect if text was written by AI. You can use either the AI Text Classifier by OpenAI or GPTZero or both of them! These are great tools that will give you a hand in determining whether specific content was generated by artificial intelligence.
It's important to note that these tools are still not 100% reliable. They will become better over time, but at the moment they're the best we got. It's important to use them as a helping hand and not take the results as completely true all of the time.
Sometimes, you'll be able to intuitively tell whether someone is using AI to generate content. This is especially the case if the person doesn't know how to write good, advanced prompts.
Unless you can tell if the content was written by AI yourself through your own methods (like comparing it to previous examples of writing from a student or employee), I suggest you use a combination of GPTZero and the AI Text Classifier to detect AI-written text.
