Study: AI Writing Models Still Exhibit Recognizable Tells
New research indicates that AI writing models, such as Opus 5.5, continue to employ distinctive word choices and sentence structures that differentiate their output from human-written text.
New research from marketing firm Graphite has examined the challenge of identifying AI-generated text. While common indicators of AI writing, like excessive use of em-dashes or the word "delve," have diminished, new, more subtle characteristics have emerged.
The study analyzed several frontier language models and found that each model exhibits unique word preferences and sentence constructions that reveal its AI origin. These "tells" can range from specific words used with much higher frequency than in human writing to particular syntactical patterns.
Graphite's analysis identified over 13,000 phrases that were at least twice as common in AI-generated content as in human-written material. A notable example is the Opus 5.5 model's tendency to use the word "dependable" significantly more often than human writers.
This research underscores the ongoing need for tools and methods to detect AI-generated content, especially as AI technologies rapidly advance and find broader application across various sectors.