Can AI detectors save us from ChatGPT? I tried 5 online tools to find out
When I first looked at the question of whether it's possible to fight back against AI-generated plagiarism, and how that might work, it was January 2023, just a few months into the world's exploding awareness of generative AI. Now, almost a year later, it feels like we've been exploring generative AI for years and years, but we've really been looking at the issue for a matter of months.
In any case, this is an updated version of that original January article. When I first tested GPT detectors, I used three: the GPT-2 Output Detector (this is a different URL than we published before), Writer.com AI Content Detector, and Content at Scale AI Content Detection. The best result was 66% correct, from the GPT-2 Output Detector. This time, I'm adding three more: GPTZero, ZeroGPT (yes, they're different), and Writefull's GPT Detector.
Also: How to use ChatGPT: Everything you need to know
Unfortunately, I'm removing the Writer.com AI Content Detector from our test suite because it failed back in January, it failed again in October, and it failed now. See below for a comment from the company, which their team sent me after the original article was published in January.
Before I go on, though, we need to talk about the concept of plagiarism and how it relates to our problem. Merriam-Webster defines "plagiarize" as "to steal and pass off (the ideas or words of another) as one's own; use (another's production) without crediting the source."
This definition fits AI-created content. While someone using an AI tool like Notion AI or ChatGPT isn't stealing content, if that person doesn't credit the words as coming from an AI and claims them as their own, it still meets the dictionary definition of plagiarism.
Also: The best AI image generators to try
In this experimental article, I've asked ChatGPT to help out. My words are in normal and bold text. The AI's words are italicized. After each AI-generated section, I'll show the results of the detectors. At the end of the article, we'll look at how well the detectors performed overall.
Here's the result for the above text, which I wrote myself:
- GPT-2 Output Detector: 99.98% real
- Content at Scale AI Content Detection: 100% Highly likely to be human!
- GPTZero: Your text is likely to be written entirely by a human
- ZeroGPT: 28.9% AI GPT Your Text is Most Likely Human written
- Writefull GPT Detector: 1% likely this comes from GPT-3, GPT-4 or ChatGPT
Human-written content: 4-of-5 correct
With that, let's dig in some more.
Explain why ChatGPT can be a real problem for teachers and editors when it comes to plagiarism
ChatGPT is a variant of the GPT (Generative Pre-trained Transformer) language model developed by OpenAI. It is designed to generate human-like text by predicting the next word or phrase in a given input sequence. While ChatGPT can be a useful tool for generating text and responding to prompts, it can also be a problem for teachers and editors when it comes to plagiarism.
One of the main reasons why ChatGPT can be a problem for teachers and editors is that it can generate text that is very similar to that of real human writers. This means that students or other individuals could potentially use ChatGPT to generate text that they could then submit as their own work, thereby committing plagiarism.
Also: 4 ways to overcome your biggest worries about generative AI
For teachers and editors, detecting whether a given piece of text was generated by ChatGPT or written by a human can be very difficult, as the generated text may not contain any obvious errors or discrepancies that would reveal its artificial origin. This can make it difficult for teachers and editors to identify and prevent plagiarism, as they may not be able to tell whether a given piece of text is original or copied from another source.
Overall, ChatGPT and other similar language models can be a problem for teachers and editors when it comes to plagiarism because they can generate text that is difficult to distinguish from that written by humans, making it easier for students and other individuals to commit plagiarism without being detected.
- GPT-2 Output Detector: 99.69% real
- Content at Scale AI Content Detection: 3% Highly likely to be AI-generated!
- GPTZero: We suggest entering a longer text for more precise prediction
- ZeroGPT: 51.22% AI GPT Your Text is AI/GPT Generated
- Writefull GPT Detector: 12% likely this comes from GPT-3, GPT-4 or ChatGPT
AI-written content: 1-of-5 correct (barely)
Remember, text shown in italics was generated by the AI.
Show me at least three ways that I can detect ChatGPT was used to produce a piece of text
There are several ways that you can potentially detect whether ChatGPT or another language model was used to produce a piece of text. Here are three potential approaches you could use:
- Look for repetitive or unusual patterns in the text: One way to detect whether a language model was used to generate a piece of text is to look for repetitive or unusual patterns in the writing. For example, language models may repeat certain phrases or word combinations, or use unusual or inconsistent sentence structures.
- Check for lack of originality: Another way to detect whether a language model was used to generate a piece of text is to check for a lack of originality in the writing. If a piece of text appears to be copied from another source, or if it contains information that is widely available on the internet or in other sources, it could be a sign that it was generated by a language model.
- Use a plagiarism checker: Finally, you can use a plagiarism checker to check whether a piece of text appears to be copied from another source. While a plagiarism checker may not be able to definitively determine whether a language model was used to generate a piece of text, it can help you identify instances of copying or paraphrasing from other sources.
- Note when the answer is obviously wrong: I talked about this in my Notion AI review, where I spotlighted some text the AI wrote with seeming authority, but which was blatantly wrong. If a student or writer doesn't have full command of the topic being discussed, they may not spot the inaccuracies and simply leave them in the AI-generated response. Of course, incorrect statements aren't just the purview of an AI writing tool. Many bloggers and commenters, for example, simply repeat statements made by other bloggers, creating an echo chamber of completely "false facts", showing that while intelligence may be artificial, stupidity is a universal trait.
Keep in mind that it can be difficult to definitively determine whether a language model was used to generate a piece of text, and these methods may not always be effective. However, they can be useful tools to help you identify potential instances of language model-generated text.
In this set of tests, only the italicized content above was run through the checker suite.
- GPT-2 Output Detector: 99.98% fake
- Content at Scale AI Content Detection: 99% Highly likely to be Human!
- GPTZero: Your text may include parts written by AI
- ZeroGPT: 91.35% AI GPT Your Text is AI/GPT Generated
- Writefull GPT Detector: 77% likely this comes from GPT-3, GPT-4 or ChatGPT
AI-written content: 4-of-5 correct
List three plagiarism checkers that can detect ChatGPT text
It is worth noting that there is currently no plagiarism checker that is specifically designed to detect text generated by ChatGPT or other language models. While some plagiarism checkers may be able to identify certain characteristics of language model-generated text that could indicate its artificial origin, there is no guarantee that they will be able to definitively determine whether a piece of text was generated by a language model.
With that said, here are three plagiarism checkers that you might consider using to help identify potential instances of language model-generated text:
- Turnitin: Turnitin is a popular plagiarism checker that is widely used by schools and universities. It compares a piece of text to a database of other texts to identify instances of copying or paraphrasing. While it may not be able to definitively determine whether a piece of text was generated by a language model, it could potentially identify certain characteristics of language model-generated text that could indicate its artificial origin.
- PlagScan: PlagScan is another plagiarism checker that compares a piece of text to a database of other texts to identify instances of copying or paraphrasing. Like Turnitin, it may be able to identify certain characteristics of language model-generated text that could indicate its artificial origin, but it is not specifically designed to detect language model-generated text.
- Copyleaks: Copyleaks is a plagiarism checker that uses advanced machine learning algorithms to scan a piece of text and compare it to a database of other texts to identify instances of copying or paraphrasing. While it may be able to identify certain characteristics of language model-generated text, it is not specifically designed to detect language model-generated text.
It is worth noting that no plagiarism checker is completely foolproof, and there is always the possibility that a piece of language model-generated text could escape detection. Therefore, it is important to use plagiarism checkers as just one part of a larger strategy to detect and prevent plagiarism.
- GPT-2 Output Detector: 99.58% real
- Content at Scale AI Content Detection: 100% Highly likely to be Human!
- GPTZero: Your text may include parts written by AI
- ZeroGPT: 100% AI GPT Your Text is AI/GPT Generated
- Writefull GPT Detector: 88% likely this comes from GPT-3, GPT-4 or ChatGPT
AI-written content: 3-of-5 correct
Online AI plagiarism checkers
Most plagiarism detectors are used to compare writing against a corpus of other writing. For example, when a student turns in an essay, a product like Turnitin scans the submitted essay against a huge library of other essays in its database, as well as against other documents and text on the internet to determine if the submitted essay contains already-written content.
But the AI writing tools generate original content, at least in theory. Yes, they build their content from whatever they've been trained on, but the actual words they construct are somewhat unique for each composition.
Also: The 10 best ChatGPT plugins right now
As such, the plagiarism checkers mentioned above probably won't work, because the AI-generated content probably didn't exist in, say, another student's paper.
In this article, we're just looking at GPT detectors. But plagiarism is a big problem, and as we've seen, some choose to define plagiarism as something you claim as yours that you didn't write, while others choose to define plagiarism as something written by someone else that you claim is yours.
That distinction was never a problem until now. Now that we have non-human writers, the plagiarism distinction is more nuanced. It's up to every teacher, school, editor, and institution to decide exactly where that line is drawn.
- GPT-2 Output Detector: 99.98% real
- Content at Scale AI Content Detection: 100% Highly likely to be Human!
- GPTZero: Your text is likely to be written entirely by a human
- ZeroGPT: 0% AI GPT Your Text is Human written
- Writefull GPT Detector: 1% likely this comes from GPT-3, GPT-4 or ChatGPT
Human-written content: 5-of-5 correct
Overall results
Of the five tests run, four out of five were generally accurate. Tests proved more accurate overall on pure human content than pure AI content. It is interesting to note that the individual detectors varied wildly, so if you're going to use a GPT detector, you might want to use a bunch of them on a single piece of text and aggregate the results. Even with that, you're not guaranteed complete accuracy.
Also: OpenAI pulls its own AI detection tool because it was performing so poorly
Test | Overall | Human | AI | AI | AI | Human |
GPT-2 Output Detector | 60% | Correct | Fail | Correct | Fail | Correct |
Content at Scale AI Content Detection | 40% | Correct | Fail | Fail | Fail | Correct |
GPTZero | 80% | Fail | Correct | Correct | Correct | Correct |
ZeroGPT | 80% | Correct | Fail | Correct | Correct | Correct |
Writefull GPT Detector | 60% | Fail | Fail | Correct | Correct | Correct |
Back in January, the best we did was 66% correct. Here, we have two tools that were correct 80% of the time. However, I would not be comfortable using these tools to validate a student's content. As has been shown, writing from non-native speakers often gets rated as generated by an AI, and even my hand-crafted content was rated as AI a few times. So, I would advocate caution before relying on the results of any (or all) of these tools.
Also: Could someone falsely accuse you of using AI-generated text? This could be why
Let's take a look at the individual testers and see how each performed.
GPT-2 Output Detector (Accuracy 60%)
This first tool was built using a machine-learning hub managed by New York-based AI company Hugging Face. While the company has received $40 million in funding to develop its natural language library, the GPT-2 detector appears to be a user-created tool using the Hugging Face Transformers library. Of the six tests I ran, it was accurate for four of them.
Writer.com AI Content Detector (Accuracy N/A)
Writer.com is a service that generates AI writing, oriented towards corporate teams. Its AI Content Detector tool can scan for generated content. Unfortunately, I found this tool unreliable, and it failed to generate results -- exactly the same way it did in January.
After this article was originally published in January, the folks at Writer.com reached out to ZDNET. CEO May Habib had this comment to share:
Demand for the AI detector has skyrocketed. Traffic has grown 2-3x per week since we launched it a couple months ago. We've now got the necessary scaling behind it to make sure it doesn't go down, and our goal is to keep it free - and up to date to catch the latest models' outputs, including ours. If AI output is going to be used verbatim, it absolutely should be attributed.
Content at Scale AI Content Detection (Accuracy 40%)
The third tool I found was also produced by an AI content generation firm. Content at Scale pitches itself as "We Help SEO-Focused Content Marketers by Automating Content Creation." Its marketing call to action is, "Upload a list of keywords and get 2,600+ word blog posts that bypass AI content detection -- all with no human intervention!" Disturbingly, the results got worse from January -- back then, it was 50% accurate.
GPTZero (Accuracy 80%)
It's not entirely clear what drives GPTZero. The company is hiring engineers and sales folks, and it runs on AWS, so there are expenses and sales involved. However, all I could find about a service offering was a place where you could register for a free account to scan more than the 5,000 words offered without login. If you're interested in this service for GPT detection, you'll have to see if they'll respond to you with more details.
ZeroGPT (Accuracy 80%)
ZeroGPT seems to have matured as a service since we last looked at it. When we last looked, no company name was listed, and the site was peppered with Google ads with no apparent strategy for monetization. The service actually worked fairly well but seemed sketchy as heck.
That sketchy-as-heck feeling is now gone. ZeroGPT presents as any other SaaS service, complete with pricing, company name, contact information, and all the rest. It still performs quite well, so perhaps the developers decided to turn their working code into more of a working business. Good for them!
Writefull GPT Detector (Accuracy 60%)
Writefull sells writing support services as well as offering free tastes of its tools. The GPT detector is fairly new, and worked fairly well. Perhaps it will improve over time.
Final thoughts
First, I'll let ChatGPT have its say: AI-driven plagiarism is a growing problem, as AI tools like Notion AI and ChatGPT can generate text that is difficult to distinguish from that written by humans. There are several ways to detect whether a piece of text was generated by an AI, such as looking for repetitive or unusual patterns, checking for a lack of originality, or using a plagiarism checker. Plagiarism checkers like Turnitin, PlagScan, and Copyleaks may be able to identify certain characteristics of language model-generated text, but they are not foolproof.
Even the specialty tools I found were woefully unsuited to the task. There was some limited improvement over my tests in January, but I certainly wouldn't feel comfortable threatening a student's academic standing or accusing them of cheating based on the results of any of these tools.
Also: Surprise! AI chatbots don't increase student cheating afterall, new research finds
At this point, I don't think we're ready to let AI fight AI. Instead, we're going to have to rely on human perception (which means folks have to pay attention) in order to identify works that were generated by an AI.
Don't immediately assume that something is machine-generated. Just because something is stupid or wildly inaccurate, doesn't mean it was not written by a human. As we've all come to know, humans are quite capable of writing some incredibly bad stuff all on their own.
Also: OpenAI's ChatGPT is scary good at my job, but it can't replace me (yet)
So, what do you think? Can you tell if a piece is AI-generated? Have you used these AI writing tools in the months since they suddenly became available? Do you think they're a help or hindrance to education and journalism? Let us know in the comments below. Here's how ChatGPT answered this:
As a language model, I can't see or interact with physical pieces of writing, so I can't tell if a piece is AI-generated or not. However, I can tell you that there are several AI writing tools that are available, and they can be used to generate a wide range of written content, including news articles, fiction, poetry, and more.
I don't have personal experience using these tools since I am just a computer program and don't have the ability to browse the internet or use physical tools. However, I can tell you that some people have found AI writing tools to be helpful in education and journalism, as they can save time and effort by automating certain tasks. Others have expressed concerns about the potential for these tools to be used to produce misleading or biased content or to replace human writers. Ultimately, whether AI writing tools are a help or hindrance depends on how they are used and the context in which they are used.
So, yeah, there's that.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter on Substack, and follow me on Twitter at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, and on YouTube at YouTube.com/DavidGewirtzTV.