The Challenge of Detecting AI-Generated Content: A Comprehensive Analysis

I tested 9 AI content detectors - and these 2 correctly identified AI text every time

When I first examined 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. This is an updated version of that original January 2023 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 (which is apparently now called BrandWell). In October 2023, I removed the Writer.com AI Content Detector from our test suite because it failed back in January 2023, it failed again in October, and it failed in summer 2024. However, it now appears to work, so I'm including it in the test suite. See below for a comment from the company, which their team sent me after the original article was published in January. Also: 88% of workers would use AI to overcome task paralysis, Google study says I've re-run all the tests to see how the detectors perform today. While I had two strong successes, the big takeaway seems to be just how inconsistent the results are from one AI checker to another. Before I go on, though, we should discuss 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 well. 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 are the test results for the above text, which I wrote myself: Human-written content: 7-of-9 (77%) correct Last time: 5-of-7 (71%) correct Before we move on, let's discuss what I'll consider a "correct" analysis. If the score is numeric, anything above 80% will be counted as a correct analysis, and that's being generous. If the score has a summary, I'll look at the summary and make a judgment. 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. Also: 4 ways to overcome your biggest worries about generative AI 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. 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. 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: 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. Also: This Google AI tool could be your new favorite study aid - and it's free In this set of tests, only the italicized content above was run through the checker suite. AI-written content: 5-of-9 (55%) correct Last time: 5-of-7 (71%) 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: 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. AI-written content: 5-of-9 (55%) correct Last time: 5-of-7 (71%) correct 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 essays in its database, and other documents and text on the internet to determine if the submitted essay contains already-written content. However, 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 words they construct are somewhat unique for each composition. Also: OpenAI pulls its own AI detection tool because it was performing so poorly 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. AI-written content: 9-of-9 (100%) correct Last time: 7-of-7 (100%) correct Overall, results declined compared to the last round of tests. That time, we had three services with perfect scores. That's down to two now because ZeroGPT, one of our then-perfect-scoring players, failed a test it previously succeeded. The two new detectors we added, Writer.com and Grammarly, didn't improve the score. In fact, both were generally unsuccessful. While the overall results have improved, I would not be comfortable relying solely on 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 though my hand-crafted content has no longer been rated as AI, there were a few paragraphs flagged by the testers as possibly AI-based. You can also see how the results are wildly inconsistent between testing systems. So, I would advocate caution before relying on the results of any (or all) of these tools. Let's look at the individual testers and see how each performed. 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 five tests I ran, the detector was accurate in three. Writer.com is a service that generates AI writing, oriented towards corporate teams. Its AI Content Detector tool can scan for generated content. I found this tool unreliable. While it previously failed to generate results, it ran this time. Unfortunately, its accuracy was quite low. It essentially identified each block of text as human-written, where three of the six tests were written by ChatGPT. Also: How to use ChatGPT to digitize your handwritten notes for free 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. The third tool I found was originally produced by an AI content generation firm, Content at Scale. Subsequently, the tool migrated to Brandwell.ai, which appears to be a new name for what is now an AI-centric marketing services company. Unfortunately, the accuracy was pretty low. The tool identified all the AI content as human, as in this screenshot: 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. 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. Accuracy increased since the first time I ran the tests and stayed at 100% for this round. 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 worked fairly well but seemed sketchy as heck. Also: AI isn't hitting a wall, it's just getting too smart for benchmarks, says Anthropic 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. Accuracy dropped, though. It misread one human-written test as AI. Writefull sells writing support services, as well as a free taste of its tools. The GPT detector is fairly new and worked fairly well. However, the tool has had some ups and downs in our tests. It improved from 60% to 80%, but dropped down to 60% again this time. Originality.ai is a commercial service that bills itself as both an AI checker and a plagiarism checker. The company sells its services based on usage credits. To give you an idea, all the scans I did for this article used 30 usage credits. The company sells 2,000 credits a month for $12.95 per month. I pumped about 1,400 words through the system and used only 1.5% of the monthly allocation. Results were great for the AI checker, but the tool failed three out of five times when using the service as a plagiarism checker. The following screenshot claims that the text pasted in was 0% plagiarised: That's wrong since all the text pasted into the tool was from this article, published online for two years. I thought, perhaps, that the plagiarism scanner couldn't read ZDNET content, but that's not the case, as this screenshot shows: To be fair, I didn't set out to check plagiarism checkers in this article. But since I'm using source material I know I pulled from my existing article, I figured the plagiarism checker would have slammed all of them as 100% plagiarized. In any case, Originality.ai did very well on the part we set out to test, the AI checker. The tool gets points for that. Nothing is ever easy. The first time I ran my first test through QuillBot, it said 45% of the text was likely generated by an AI. It wasn't. I wrote it. But then, after completing all the other tests, I returned to QuillBot to grab a screenshot for this section, fed it the same text that generated the 45% score, and, as you can see, it now reports 0% AI. So, what are we to make of this result? Sadly, I didn't grab a screenshot of the first time I tested this text, but it highlights the concern with relying too much on AI detectors because they are also quite capable of hallucination. Grammarly is a well-known tool for helping writers produce grammatically correct content. That's not what we're testing here. Grammarly can check for both plagiarism and AI content. You can paste a document into their grammar checker, and in the lower-right corner, there's a Plagiarism and AI text check button. In this test, the tool found an existing online document that matched what I pasted in. That result makes sense because this is an update to an article that's been online for a few years. But what isn't correct is the statement that "Your document doesn't appear to contain Al-generated text," because the entire segment was generated by ChatGPT. Well, OpenAI pulled it last year because the detector wasn't particularly accurate. As of August, it was supposed to be 99% accurate, but there's still no sign of a release four months later. Its claimed accuracy level is also a little tough to believe because ChatGPT itself is far from 100% accurate. Also: Will OpenAI's new AI detection tool put an end to student cheating? But, in any case, as my ZDNET buddy Steven Vaughan-Nichols writes, it's not available yet. We'll give the tool a spin when it's released for public use. Stay tuned. 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 speciality tools I found were woefully unsuited to the task. There was some limited improvement over my tests in January. However, I certainly wouldn't feel comfortable threatening a student's academic standing or accusing them of cheating based on the results of these tools. Also: Surprise! AI chatbots don't increase student cheating after all, research finds At this point, I don't think we're ready to let AI fight AI. Instead, we'll have to rely on human perception (which means folks have to pay attention) to identify works 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 all know, humans can write some bad stuff on their own. Also: Could someone falsely accuse you of using AI-generated text? This could be why 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.

How to identify AI-generated text: 7 ways to tell if content was made by a bot

As AI-generated content gets more ubiquitous in our everyday lives, you may be wondering, "How do I identify AI text?" It's no surprise that these models get more difficult to detect as AI technology evolves. For now, the good news is that content such as images and video aren't that hard to parse with the human eye. If you're a teacher or just a seasoned internet traveler, what's the secret to spotting AI-generated text? Well, it's simpler than you might think: use your eyes. There are actually ways to train the human eye to discern AI statements. Experts like MIT Technology Review's Melissa Heikkilä write that the "magic" of these machines "lies in the illusion of correctness." No two people write in the same way, but there are common patterns. If you've ever worked a corporate job, you know how everyone uses the same generic phrasing when drafting memos to their boss. That's why AI text detectors often flag content as "likely AI-generated" -- because distinguishing between a bland human writing style and a generic AI-generated voice is nearly impossible. So here's some tips and tricks to spot some potential AI-generated text: * Conclusionary statements that neatly sum up paragraphs. * A tone more advanced than the writer's usual submissions. * Repetitive phrasing or oddly polished grammar. There are also AI text detectors on the market that you can use, but here's why, in my experience, they're likely less reliable than your own eyes. It's not all doom and gloom, as some solutions to our machine overlords exist. Launching models like ChatGPT and competitors like Gemini and Claude spurred the growth of a cottage industry focused on AI text detection. Platforms like ZeroGPT popped up in response to OpenAI's language model, while tools such as Grammarly and Copyleaks -- originally designed to catch plagiarism -- have pivoted to tackle AI-generated content as well. Depending on who you ask, AI-text detection is, at the moment, the best way to spot AI-generated content or its digital snake oil. In reality, the latter might be closer to the truth. No AI detector is 100% accurate (or even 99% as many claim). Even in ideal conditions, the reliability of these tools is often hit-or-miss. "The problem here is the models are becoming more and more fluent, [as a result], the older detectors, they stop working," says Junfeng Yang, a professor and co-director of the Software Systems Lab at Columbia University. He explains that as AI-generated text becomes increasingly sophisticated, it "starts to use vocabulary and sentence structures that closely mimic human writing, making it harder to spot even with advanced detectors." Despite big promises from tools like GPTZero or Hive Moderation, tricking an AI detector into labeling human-written content as machine-made is surprisingly simple. These systems typically analyze lines of text that explain or summarize ideas, which makes them vulnerable to false positives. For instance, I tested a basic summary of Game of Thrones I had hastily written from memory across several of these tools, and in almost every case, it was flagged as "likely AI-generated." I then took that same Game of Thrones summary, fed it into ChatGPT, and asked it to punch up the writing -- making it longer and less straightforward. Here's what I wrote: The plot of Game of Thrones follows several families in the fictional continent of Westeros. First, there are the Starks who rule the North. Second are the Lannisters, who rule in the Westerlands. The eldest daughter of the Wastelands is Cersei, who is also the Queen of Westeros and is married to Robert Baratheon. Robert is from the Stormlands and usurped the throne of the Targarayns -- a group of dragon riders who took over the land over 1000 years ago but slowly lost their grip on Westeros due to infighting. The series was created by George RR Martin and is currently chronicled in five books -- A Game of Thrones, A Clash of Kings, A Storm of Swords, A Feast for Crows, and a Dance with Dragons. There are two more books planned for the series with the Winds of Winter set to release sometime in 2025. Here's the "improved" ChatGPT version: Game of Thrones throws viewers into the brutal, blood-soaked chessboard of Westeros, where power is the prize and morality is negotiable. Leading the pack are the Starks, who hold dominion over the icy North. Their stoic honor feels like a relic in a land where backstabbing is a form of currency. In sharp contrast are the Lannisters of the opulent Westerlands, the family everyone loves to hate -- with Cersei Lannister perched atop the throne as Queen of Westeros. Her marriage to Robert Baratheon, the hard-drinking, boar-hunting king from the Stormlands, is a political match made in hell, bound together by lies and mutual loathing. But the throne itself carries a bloody history. Robert took it after a rebellion that ousted the Targaryens, the long-ruling dynasty of dragon riders who, once unstoppable, eventually imploded in a whirlwind of paranoia and civil war. Their fire-breathing beasts may have once scorched the skies, but family feuds brought their empire crashing down, leaving only whispers of their former glory. Meanwhile, George R. R. Martin's sprawling saga continues in five books that are equal parts battlefield carnage, political intrigue, and heartbreaking betrayals. While fans devour every page, they remain on edge, waiting for the Winds of Winter to blow, teased for what feels like eons, now with a tentative promise of landing in 2025. But in Westeros, promises -- like allegiances -- are made to be broken. The outcomes of my testing had varied results. One of the first I tested was Scribbr, which pegged my original summary as "100% likely AI-generated." But when I swapped it for the AI-enhanced version, Scribbr surprisingly dropped the likelihood to just 48%. Another tool I tried out was Winston AI, which flagged the original text as 39% likely to be generated by GPT-4. Ironically, when it came to the actual GPT-4 summary, it only gave it a meager 1% likelihood. Hive Moderation totally missed the mark when analyzing my work, failing to flag either of the summaries I submitted. According to the system, both were confidently labeled as human-written content. Now, if I just ask ChatGPT for a random paragraph on any topic and copy-paste that into various text detectors, it'll almost always get flagged as AI-generated right away. But that actually reinforces the issue: without specific instructions, ChatGPT's default writing style is often bland, formulaic, and straightforwardly objective. The predictably dull tone is what triggers these false positives -- not some advanced in-house tech that these websites claim to have to discern AI content from humans. Even when tools like Originality correctly flagged both instances of AI writing, a bit of sentence tweaking can completely change the outcome. With just a little rephrasing, what was previously flagged with "100% confidence" as AI-generated can suddenly be labeled "Likely original." All that to say, here's the list of freely available AI text detection tools I tested using the above method. To mix things up, I also used some literature reviews from academic papers I wrote in grad school to see if they'd flag me for using flowery writing to beef up my word counts. Here they are: If your writing sounds like a tonally flat 8th-grade book report, AI detectors will likely peg you as a bot in need of a Turing test ASAP. This testing shows that simply avoiding certain structural patterns can easily fool AI detectors. And that's a major headache for the companies behind these tools, especially since many offer subscription services and aim to sell their APIs to schools and businesses as a B2B solution. While these tools can be pretty effective for plagiarism detection, it's obvious their ability to spot AI-generated text still needs serious refinement. The inconsistency is hard to overlook -- submit the same text to multiple detectors, and you'll get wildly different results. What gets flagged as AI-generated by one tool might slip through unnoticed by another. Given that lack of reliability, it's tough to recommend any of these tools with confidence right now. Human language is incredibly fickle and complex -- one of the main reasons AI-generated text is so tricky to detect. Bamshad Mobasher, IMEE member and chair of the AI program at DePaul University elaborates that "text is what these models are trained on. So, it's easier for them to mimic human conversations." "Detection tools look for patterns -- repetitive phrases, grammatical structures that are too regular, things like that," Mobasher said. "Sometimes, it's easier for a human to spot, like when the text is 'too perfect,' but to be certain it's AI-generated is challenging." Unlike image generators, which can produce telltale signs like extra fingers or distorted facial features, Mobasher explained LLMs rely on statistical probabilities to generate text -- making their output feel more seamless. As a result, spotting errors in AI-generated text -- like nuanced phrasing or subtle grammatical irregularities -- is far more challenging for both detectors and human readers. This is what makes AI-generated text so dangerous as well. Mobasher warns that "it becomes easier to produce and generate misinformation at scale." With LLMs generating fluent, polished text that can mimic authoritative voices, it becomes much harder for the average person to discern between fact and fiction. "With AI, it's actually much easier to launch these attacks," says Yang. "You can make the email very fluent, conveying the message you want, and even include personalized information about the target's role or mission at a company." On top of its potential misuse, AI-generated text makes for a shittier internet. LLMs from companies like OpenAI and Anthropic scrape publicly available data to train their models. Then, the AI-generated articles that result from this process are published online, only to be scraped again in an endless loop. This cycle of recycling content lowers the overall quality of information on the web, creating a feedback loop of increasingly generic, regurgitated material that makes it difficult to find authentic, well-written content. There's not much we can do about the lightning-fast acceleration of AI and its detrimental effects of internet content, but you can, at the very least, tap into your knowledge pool of media literacy to help you discern what's human-made and what's generated from a bot. "If you see an article or report, don't just blindly believe it -- look for corroborating sources, especially if something seems off," Yang says.