New Study Reveals Persistent Writing Quirks Across Leading LLMs

Marketing research firm Graphite examined 10,000 pre‑ChatGPT articles and asked several state‑of‑the‑art models to rewrite them from summaries. By comparing phrase frequency between human and machine output, the team identified more than 13,000 expressions that appear at least twice as often in AI‑generated text.

Claude Opus 5.5’s strongest marker is the word “dependable,” which occurs 23 times more often than in human writing. The model also favors the construction “is more than an X, it’s a Y” and repeatedly signals importance with “this matters” (116×) and “why X matters” (92×). OpenAI’s Astra, by contrast, leans on “another dimension” and hedges with “may provide” or “can provide,” while its signature corrective framing — “not simply X” or “rather than relying on X” — appears over 100 times more often than in human prose.

All three labs have dramatically reduced em‑dash usage: Opus 5.5 cuts them by 99 percent versus its predecessor, Astra by 88 percent versus humans, and Gemini 3.1 Pro almost eliminates them entirely. Yet the overall count of distinctive tells remains stable, suggesting that removing well‑known quirks simply uncovers new ones.

Why it matters for GPU and AI infrastructure

Detecting and benchmarking these linguistic fingerprints requires massive inference runs across diverse models, a workload that scales directly with GPU capacity. Cloud providers that offer flexible, high‑throughput GPU clusters enable researchers to iterate quickly on evaluation pipelines, accelerate model‑comparison studies, and ultimately deliver more transparent, trustworthy AI services.