Benchmark finds AI rewrites overuse 'though' 5x more than humans: 0.89 vs 0.18 per 1k words
rseroter · x · 2026-11-17
A benchmark by kweinmeister tests seven agent skills designed to strip telltale signs of AI writing. Data across 150 AI-rewritten drafts vs. 50 pre-LLM human passages shows AI rewrites drastically overuse ordinary connectors and modifiers: "though" at 0.89 vs 0.18 per 1,000 words, "specifically" 0.64 vs 0.12, "furthermore" 0.51 vs 0.00, "essential" 0.49 vs 0.06, with "vital" and "alongside" at 0 in human writing. These words now serve as fingerprints for detecting AI-generated text.
More from Models
- User says OpenAI refused $570 refund for auto-renewal with no explanation — OverLordGoldDragon · 2026-11-17
- Gemini spots the trick in a handwritten integral troll that stumps ChatGPT — vedantmisra · 2026-11-17
- GPT-6 Astra predicted its own failure mode in Codenames — then stepped right into it — Hardicecream17 · 2026-11-17
- Tencent Hy4 Flash scores 5.74 on RuneBench at just $0.03 per run, cheapest on Pareto frontier — mariofilhoml · 2026-11-17
- Tencent Hunyuan Hy4 Flash hits 5.74 on RuneBench at ~$0.03 per run, cheapest on Pareto frontier — mariofilhoml · 2026-11-17
- Open-source models hit 30% enterprise token share in 5 months, up from near zero — pwendell · 2026-11-17