6 Best A.I. Content Detectors To Identify GPT Text
So A.I. writing tools are amazing. We know that and they can be super helpful to come up with ideas, even to help you write faster.
However, there are multiple occasions in which A.I. content is just not ideal.
The most obvious situation is a school assignment. Back in my school years, teachers could notice when a not-so-smart kid was using Google Translate for their Foreign Language homework.
Now, students might be tempted to use Jasper or other A.I. writing tools to create A.I. content and forget about spending time and effort on those long essays. Although an experienced teacher might notice the cheating, it can be hard for others.
You might think your students are too young or not as tech-savvy to be using A.I. writing tools, but I assure you they are not. This type of technology is super easy to use and they might already be giving you assignments with A.I. content.
But if you’re a teacher or anyone doubting the content you’re given is legit or not, don’t worry. With A.I. content, comes A.I. content detection tools too.
In this list, we’ll give you the best A.I. content detection tools so you can spot those liars and don’t let them fool you.
Best A.I. Content Detection Tools
To make this more fun, I took an old post I wrote in 2021. It was about Shortly.AI and I used the A.I. to see how it worked. Pieces of the post included A.I. content.
I copied the post and scan it with each A.I. content detection tool in this list to test them. Let’s see the results.
Note: if you want to join me in this experiment, make sure the content you want to analyze isn’t too short. Most tools can detect A.I. content only if the input is 100 words or more.
OriginalityAI is a plagiarism checker and A.I. content detector built for content publishers. It has multiple features like Team Management, Scan History by User and Website Scan.
According to a recent study, OriginalityAI can detect content produced by GPT 3.5 (the new model for AI-generative text) with the same accuracy as it can for GPT-3, which is 94%.
I tested OriginalityAI with its Chrome extension and at first found it a bit confusing. OriginalityAI, unlike the Paraphrasing tool and others, requires you to sign up to use it and it delivers all the results on its dashboard, even if you’re using the Chrome extension.
Then I figured it out and I was able to scan content on Google Docs and practically anywhere online. As you can see, the tool was able to detect my content contained some A.I. in it.
Although it’s a paid tool and you need to pay $0.01 per 100 words per scan, it does give you some free credits so you can get a sense of how it works.
Each content scan will give you an AI detection score as well as a plagiarism score.
Originality is still in Beta and sadly, it currently only supports the English language.
But the good part is it comes with API access.
GPTZero is one of my favorites and it was developed by a 22-year-old undergraduate student!
Edward Tian, who studies computer science and journalism at Princeton University, created this free, minimalistic, easy-to-use tool to detect machine-generated text.
GPTZero analyzes content that’s between 250 words and 500 words. You can paste the input directly on their website or add a file. You can also upload multiple files at once if you have an entire classroom.
The results are instant and quite detailed. In my case, GPTZero effectively detected there were parts written by an A.I.
It also highlighted which parts were more likely to be written by an A.I. and gave me an average perplexity and the sentence with the highest perplexity.
Perplexity refers to how complex or random a text might be. Naturally, the more complex or random, the more human it is.
There’s also a burstiness indicator which is a measurement of the variation in perplexity. According to Tian, machines have pretty constant writing over time, contrary to humans which might have sudden bursts of creativity.
This detection tool is still quite new so they have a product waitlist you can join or, if you prefer, there’s already an API available.
I especially like GPTZero because you can use it for free, online, without even registering. Also, although I couldn’t find documentation talking about the languages it supports, I tried Spanish, English and Russian and it successfully analyzed them all!
Giant Language model Test Room
This A.I. content detection tool is much more for language nerds like me than for teachers or bosses searching for direct answers. It was developed by the MIT-IBM Watson A.I. Laboratory, Harvard NLP and researchers.
And although it was made back in 2019 to test against GPT-2 text (and not ChatGPT models), it’s still very interesting and fun to use.
The GLTR demo enables forensic inspection of the visual footprint of a language model on the input text.
A word that ranks within the most likely words is highlighted in green (top 10), yellow (top 100), red (top 1,000), and the rest of the words in purple. The fraction of red and purple words represent unlikely predictions and it increases when you move to the human-created texts.
So even though it’s not super updated, it provides a complete scan and takes into account the nuance of the text.
And if we talk about GLTR, we have to talk about OpenAI. One of the developers of GLTR pointed out OpenAI as the most reliable ChatGPT detector.
They have trained, after all, the ChatGPT model.
To use the tool, you need to create an account, but don’t worry, it’s free.
The page clarifies that the classifier was trained in English, which is expected. But it also says something unexpected and interesting to note: it was also trained with content written by adults so it’s likely to get things wrong if the text was written by children.
The detector doesn’t have a word limit so I was able to paste my entire post. It labeled it, however, as very unlikely A.I. generated. Each document is labeled as either very unlikely, unlikely, unclear if it is, possibly or likely AI-generated.
I then tried with a text generated by OpenAI itself using ChatGPT and the result was possible A.I. generated. So chances are it detects A.I. better if it’s the ChatGPT model.
Copyleaks is a plagiarism detector software. Their A.I. content detector is still in Beta but they assure it has a 99.12% accuracy.
As you can see, you can use it online for free. Once you copy the text, it will tell you whether the content is A.I. or not and its probability percentage.
Unlike the previous tools, Copyleaks won’t differentiate between different parts of the text. It will mark it all as A.I. or human, with no nuance.
I have to say, Copyleaks had some trouble with my Shortly post but to be fair, it was a hard test because most of the content was indeed human. I then tried it with a piece that was 100% A.I. content and it successfully detected it.
The platform can also detect A.I. across different languages, including English, Spanish, German, French and Portuguese, with more languages to come. Naturally, as always, it works best with English, but the accuracy of other languages is not far behind and getting more and more accurate by the day.
Another thing worth notice is they are committed to staying updated and will support ChatGPT 4 once it is released!
Finally, Copyleaks is very versatile, it has an API integration and a Chrome extension. Moreover, there’s also an LMS integration so you can detect A.I. content right from your LMS platform.
The Paraphrasing tool offers an A.I. article generator and an A.I. content detector. The detector has a 90% accuracy, although I have to say it didn’t spot one of the inputs I summited.
What makes it different from other tools in this list is that it also has an A.I. content bypasser.
That’s right, as well as there are content detection tools out there, there are also tools to bypass that detection. So if in this case, you’re a writer that wants to make sure you’re not getting caught, you can use this platform.
First, you’ll use the A.I. content detection tool to see whether you’re being too obvious or not. If the answer is yes, you can then go to the A.I. content bypasser tool, copy the same text and generate a new one.
I tried it and then came back to the detection tool and although it certainly improved, it just wasn’t enough. The human content percentage increased to 15% and the A.I. was 85% which is still most of the text.
I will suggest using the Paraphrasing tool with some caution.
If you’re the person using A.I. writing tools, be aware people have their ways to detect it. So keep that in mind the next time you sit down to write for a client or a school assignment. You might get caught!
But it’s not even about your superiors noticing. Google —maybe our true superior—can notice too and search engines don’t like content that doesn’t read as human.
The challenge of finding a good A.I. content detector is real. Mainly because A.I. writing tools are working with very recent models, the GPT-3 and ChatGPT, so these detectors need to evolve with them.
You’ll find plenty of tools that detect GPT-2 generations, with high chances of leaving your GPT-3 content undetectable.
All in all, in my opinion, the best A.I. content detection tool is GPT-Zero. It’s the most accurate and it’s also very user-friendly. Not to mention it’s free!
Have you tried any of these tools?
Do you have any anecdotes of being caught up cheating or you catching someone else?
We’ll love to read some funny stories.
Tell us in the comments!