AI Models Exposed by Favorite Phrases

🔥 What happened Researchers at Graphite Growth identified 13,000 phrases that AI models use at least twice as often as humans. Claude Opus 5.5 uses “reliable” 23 times more frequently, while GPT-6 Astra favors corrective formulations – each about 100 times more often than human texts. 💡 Why it matters The em-dash is nearly extinct, but the total number of detection markers remains stable. Each model version has its own favorite phrases – an arms race that never ends. LinkedIn already introduced a reporting feature for AI-generated posts. ⚡ Our take Labs can only control so much – with billions of parameters, new patterns slip through. To detect AI text, don't fixate on individual words; look at the frequency.