Inside weeks of ChatGPT’s launch, there have been fears that students can be using the chatbot to spin up passable essays in seconds. In response to those fears, startups began making products that promise to identify whether text was written by a human or a machine. 

The issue is that it’s relatively easy to trick these tools and avoid detection, in line with recent research that has not yet been peer reviewed. 

Debora Weber-Wulff, a professor of media and computing on the University of Applied Sciences, HTW Berlin, worked with a bunch of researchers from a wide range of universities to evaluate the power of 14 tools, including Turnitin, GPT Zero, and Compilatio, to detect text written by OpenAI’s ChatGPT. 

Most of those tools work by on the lookout for hallmarks of AI-generated text, including repetition, after which calculating the likelihood that the text was generated by AI. However the team found that every one those tested struggled to choose up ChatGPT-generated text that had been barely rearranged by humans and obfuscated by a paraphrasing tool, suggesting that every one students must do is barely adapt the essays the AI generates to get past the detectors. 

“These tools don’t work,” says Weber-Wulff. “They don’t do what they are saying they do. They’re not detectors of AI.”

The researchers assessed the tools by writing short undergraduate-level essays on a wide range of subjects, including civil engineering, computer science, economics, history, linguistics, and literature. They wrote the essays themselves to make sure the text wasn’t already online, which might have meant it’d have already got been used to coach ChatGPT.

Then each researcher wrote an extra text in Bosnian, Czech, German, Latvian, Slovak, Spanish, or Swedish. Those texts were passed through either the AI translation tool DeepL or Google Translate to translate them into English. 

The team then used ChatGPT to generate two additional texts each, which they barely tweaked in an effort to cover that it’d been AI-generated. One set was edited manually by the researchers, who reordered sentences and exchanged words, while one other was rewritten using an AI paraphrasing tool called Quillbot. In the long run, they’d 54 documents to check the detection tools on.

They found that while the tools were good at identifying text written by a human (with 96% accuracy, on average), they fared more poorly when it got here to spotting AI-generated text, especially when it had been edited. Although the tools identified ChatGPT text with 74% accuracy, this fell to 42% when the ChatGPT-generated text had been tweaked barely.

These sorts of studies also highlight how outdated universities’ current methods for assessing student work are, says Vitomir Kovanović, a senior lecturer who builds machine-learning and AI models on the University of South Australia, who was not involved within the project.

Daphne Ippolito, a senior research scientist at Google specializing in natural-language generation, who also didn’t work on the project, raises one other concern.

“If automatic detection systems are to be employed in education settings, it’s crucial to grasp their rates of false positives, as incorrectly accusing a student of cheating can have dire consequences for his or her academic profession,” she says. “The false-negative rate can also be essential, because if too many AI-generated texts pass as human written, the detection system isn’t useful.” 

Compilatio, which makes one in all the tools tested by the researchers, says it’s important to do not forget that its system just indicates suspect passages, which it classifies as potential plagiarism or content potentially generated by AI.

“It’s as much as the faculties and teachers who mark the documents analyzed to validate or impute the knowledge actually acquired by the creator of the document, for instance by setting up additional technique of investigation—oral questioning, additional questions in a controlled classroom environment, etc.,” a Compilatio spokesperson said.

“In this fashion, Compilatio tools are a part of a real teaching approach that encourages learning about good research, writing, and citation practices. Compilatio software is a correction aid, not a corrector,” the spokesperson added. Turnitin and GPT Zero didn’t immediately reply to a request for comment.

We’ve known for a while that tools meant to detect AI-written text don’t at all times work the best way they’re purported to. Earlier this 12 months, OpenAI unveiled a tool designed to detect text produced by ChatGPT, admitting that it flagged only 26% of AI-written text as “likely AI-written.” OpenAI pointed MIT Technology Review towards a piece on its website for educator considerations, which warns that tools designed to detect AI-generated content are “removed from foolproof.”

Nonetheless, such failures haven’t stopped firms from rushing out products that promise to do the job, says Tom Goldstein, an assistant professor on the University of Maryland, who was not involved within the research. 

“A lot of them will not be highly accurate, but they will not be all a whole disaster either,” he adds, declaring that Turnitin managed to realize some detection accuracy with a reasonably low false-positive rate. And while studies that shine a lightweight on the shortcomings of so-called AI-text detection systems are very essential, it will have been helpful to expand the study’s remit to AI tools beyond ChatGPT, says Sasha Luccioni, a researcher at AI startup Hugging Face.

For Kovanović, the entire idea of attempting to spot AI-written text is flawed.

“Don’t attempt to detect AI—make it in order that the usage of AI isn’t the issue,” he says.