AI detection tools face accuracy challenges in identifying human writing

Discussions regarding the reliability of artificial intelligence detection tools highlight significant inaccuracies in identifying AI-generated text. Testing has shown that even human-written content can trigger high AI-probability scores from various detectors. For instance, one human-authored blog post received a 94 percent likelihood of being AI-generated from GPTZero, while Originality.ai rated it at 12 percent and Copyleaks at 61 percent.

Common patterns that humans use to identify AI writing include repetitive sentence lengths, an overreliance on lists, the frequent use of specific words like “delve” or “unlock,” and a lack of personal opinion. However, these are merely patterns rather than definitive proof. There is growing concern regarding “AI slop” and the use of generative AI in historical or creative contexts, such as modified vintage advertisements, which can create subtle, unsettling inaccuracies.