Graphite Study Finds AI Models Still Leave Thousands of Writing Tells
aiinsider ·
A recent study by marketing firm Graphite reveals that frontier AI models continue to exhibit thousands of distinctive writing patterns that distinguish them from human authors. Researchers identified approximately 13,000 phrases that appear at least twice as frequently in AI-generated text compared to human prose. To conduct this analysis, the team used 10,000 pre-ChatGPT articles as a control group and had various models rewrite summaries of these pieces. The findings highlight specific stylistic quirks, such as Claude Opus 5.5’s heavy reliance on the word "dependable" and its tendency to explain significance up to 116 times more often than humans. Meanwhile, OpenAI’s Astra model frequently employs hedging language and corrective framing, constructions that are over 100 times more common in its output than in human writing. Although labs have successfully reduced the use of em-dashes, Graphite’s chief AI officer notes that new tells emerge with each version, keeping the total number of identifiable markers steady. The study suggests that while some models like Claude are aligning closer to human patterns, others like GPT are drifting further away, indicating limited lab control over these inherent quirks.
aiinsider