The rise of AI tools like ChatGPT has made it easy for anyone to generate text that looks like human writing. While much of this is harmless, schools are dealing with plagiarism, and bad actors are using these tools to spread misinformation.

Now, researchers at Johns Hopkins University have developed a tool that can do more than just detect AI-generated writing. It can also identify which specific AI model created the text. This breakthrough could help catch cheaters and track the sources of fake news.

“We’ve shown for the first time that AI writing has unique features, just like human writing,” the researchers explain. “This lets us reliably detect machine-generated text and figure out which model produced it.”

The tool, trained on anonymous Reddit posts, works in any language and is freely available for anyone to use. It has already been downloaded 10,000 times. While other AI detectors exist, this one stands out for being highly accurate and quick to adapt to new AI systems.

Writing styles

The team didn’t set out to detect AI writing. Instead, their goal was to study how to tell different human writers apart. But they found that the same techniques work well for spotting machine-generated content.

The discovery that each AI writing tool has its own style surprised the researchers, who had assumed all machine-generated text would look similar.

The system works by using neural networks to spot subtle patterns in text. Unlike traditional methods, which focus on specific features like punctuation or grammar, the AI decides for itself what to analyze. This approach builds on techniques used by law enforcement to match writing to suspects but automates the process and scales it up.

The stakes are high. While many uses of AI writing are harmless, some are not. Schools are grappling with the rise of AI-assisted cheating, and disinformation campaigns are becoming harder to trace. This tool could provide a way to navigate the challenges of an AI-driven world.