New Method Catches AI-Written Text and Hidden Data Use

When you read something online, how do you know a human actually wrote it? AI models now produce text that sounds natural, which makes it harder to spot fake reviews, spam, or misinformation. This new research tackles that problem with a smarter detection method called likelihood-array regression, or LAR.

Instead of just checking how likely a full sentence is for an AI, LAR examines each word under different amounts of context — like reading a sentence with the first few words, then more, then all of it. That creates a detailed “array” of information. The researchers found this approach catches AI-written text more accurately than older likelihood-based methods. It also improves membership inference, which is a way of asking: “Was this document part of the AI’s training data?” That matters for privacy and copyright.

The results are promising, but there are honest limits. The method works best when the AI model is known; in real life, you don’t always know which model produced a text. Also, as AI keeps improving, detection tools need constant updates. LAR is not yet a commercial product, so it doesn’t directly protect you today — but it shows that keeping punctuation, phrasing, and word-choice will become increasingly hard to fake. Expect this kind of tool to appear in content filters, plagiarism checkers, and AI safety products soon.