牛津實測：AI 改寫貼文會悄悄帶偏立場把簡體中文原文的重點轉成繁體中文讀者更容易閱讀的脈絡。核心在於 Oxford researchers warn that AI-mediated rewriting can shift user opinions subtly even when asked to preserve meaning.
變化在哪裡
目前可核驗的事實包括：topics including gun control, cannabis legalization, atheism and the death penalty, models drifting in consistent directions, simulations using X and Facebook network data, Grok post-explanation samples on abortion, and ICML Seoul AI4Good plus technical AI governance workshop acceptance。這些細節讓新聞不只停留在發布會口號。
為什麼重要
The risk is not only deepfakes or obvious propaganda; it can sit inside everyday writing assistants and rewrite buttons. AI 新聞已經不能只看模型分數，還要看成本、供應、監管與真實工作流程。
接下來看什麼
Platforms need to disclose and test how rewriting features change stance, not just whether generated content is labeled. 下一步要看的不是更多宣傳，而是第三方測試、客戶續用與長時間穩定性。
參考來源：Mirage News and University of Oxford / OII paper AI-Mediated Communication Can Steer Collective Opinion、CocoLoop。