The best guide to spotting AI writing comes from Wikipedia

It’s no secret that many of us have experienced a nagging feeling that the text we’re reading might have originated from an artificial intelligence. However, pinpointing this can be quite challenging. Last year, there was a surge of speculation that certain words, like "delve" or "underscore," could serve as giveaways for AI-generated content. Yet, as language models have evolved, identifying these markers has become increasingly difficult. Fortunately, Wikipedia editors have taken strides in addressing this issue through their initiative known as Project AI Cleanup. This project aims to manage the flood of AI submissions, which has become a pressing concern given the sheer volume of edits made on the platform daily. The result is a comprehensive guide titled "Signs of AI Writing," which stands out as an invaluable resource for discerning AI-generated text from human-written prose. What sets this guide apart is its focus on specific language patterns and phrases that are relatively uncommon on Wikipedia but prevalent across the broader internet, which informs the training data for AI models. For instance, the guide notes that AI-generated submissions often stress the importance of the subject in vague and generic terms, such as referring to a “pivotal moment” or “broader movement.” Moreover, these entries frequently highlight minor media references in a way that resembles personal biographies rather than objective reports. An intriguing feature identified in the guide is the use of ambiguous phrases that claim significance, such as stating that something is “emphasizing the significance” or “reflecting the continued relevance” of a concept. This grammatical structure, known as the “present participle,” is subtle yet pervasive once recognized. Additionally, the guide points out a tendency toward ambiguous marketing language often found online. Descriptions tend to be overly positive, with landscapes described as "scenic" and views labeled as "breathtaking," creating a tone reminiscent of television advertisements. Overall, while some may have believed that the rapid development of language models would make them untraceable, the patterns highlighted in this guide reveal that these traits are deeply ingrained in their training. Though they may be masked, completely eliminating them is unlikely. As public awareness of these characteristics grows, the implications for communication and authenticity could be profound.