Hustler Words – As large language models (LLMs) become increasingly integrated into our daily digital lives, a new arms race has emerged: the quest to distinguish human creativity from machine-generated prose. While the early "red flags" of AI writing—such as the excessive use of em-dashes or the ubiquitous verb "delve"—are fading, a groundbreaking new study suggests that AI still leaves a distinct linguistic fingerprint.
A recent deep dive by the marketing intelligence firm Graphite has uncovered a massive catalog of "tells"—specific linguistic patterns and vocabulary choices that betray an AI’s origin. The researchers identified a staggering 13,000 phrases that appear at least twice as frequently in AI-generated text compared to human writing.
The study utilized a rigorous methodology, comparing a control group of 10,000 human-written articles published pre-ChatGPT against content rewritten by various frontier models. This allowed researchers to isolate the unique "personality" of each AI.
The findings reveal a fascinating divergence in how different AI families evolve. According to Greg Druck, Chief AI Officer at Graphite, Anthropic’s Claude models are actually trending toward more natural, human-like word distributions. Conversely, OpenAI’s GPT models appear to be drifting further away from human linguistic norms.
Take Claude Opus 5.5, for example. While it has successfully moved away from certain predictable sentence structures, it has developed its own eccentricities. The model has a peculiar obsession with the word "dependable," using it 23 times more often than a human would. Perhaps even more telling is its tendency to lecture the reader; the phrase "this matters" appears 116 times more frequently in Opus 5.5’s output than in human prose.
OpenAI’s Astra presents a different profile. It frequently attempts to add "another dimension" to topics and relies heavily on "corrective framing"—a structural habit where the model defines a subject by what it is not rather than what it is. This specific construction was found to be 100 times more common in Astra’s writing than in human-authored text.
Interestingly, the industry has successfully addressed some of the most obvious giveaways. The heavy-handed use of em-dashes has plummeted across the board; Gemini 3.1 Pro has nearly eradicated the punctuation mark entirely, and Opus 5.5 has seen a 99% reduction compared to its predecessor.
However, Druck remains cautious about the possibility of ever truly "humanizing" AI. Despite the marketing claims from labs like Anthropic and OpenAI regarding improved clarity and natural flow, the underlying patterns persist.
"These are massive models with billions of parameters," Druck noted in an interview with hustlerwords.com. "They have a finite number of tests they can run, and things inevitably slip through." As AI continues to evolve, the battle between sophisticated generation and human detection is far from over.
Leave a Comment