AI writing patterns persist despite efforts to mimic humans

A new study by Graphite reveals that AI-generated writing still exhibits distinct patterns, despite efforts to make it more human-like. The research analyzed 10,000 articles written before ChatGPT's release, comparing them to AI-generated content. Claude Opus 5.5 frequently uses the phrase 'this matters' 116 times more than human writing, while OpenAI's Astra relies on 'corrective framing,' where topics are defined as 'not simply X.' Graphite identified 13,000 phrases that appear twice as often in AI content as in human writing. Greg Druck, Graphite’s chief AI officer, noted that while models like Opus 5.5 have reduced em-dash usage by 99% compared to earlier versions, other patterns persist. Druck expressed skepticism about labs' ability to fully eliminate these telltale signs, citing the complexity of large models with billions of parameters. Meanwhile, OpenAI and Anthropic have made claims about improving clarity and reducing jargon in their models. However, Graphite’s findings suggest that while some tells are being addressed, new ones continue to emerge, with each model version displaying its own unique quirks.