Recent academic studies and company reports show that large language models (LLMs) are changing how we communicate online. This is happening because these AI tools often introduce political bias and increase the amount of automated spam.

近期的學術研究與公司報告顯示，大型語言模型 (LLM) 正在改變我們在網路上的溝通方式。這是因為這些 AI 工具經常引入政治偏見，並增加了自動化垃圾訊息的數量。

Research from the Oxford Internet Institute and the Hasso Plattner Institute shows that AI tools used for writing and summarizing often change the original meaning of a text to match specific political views. For example, models from Meta, Google, Alibaba, and Mistral tended to show liberal views on topics like climate change and feminism. In contrast, xAI's Grok showed a right-wing perspective, especially regarding abortion. These changes happened even when the AI was told to keep the original meaning, which suggests that these models have built-in political leanings.

牛津網路研究院與 Hasso Plattner 研究院的研究顯示，用於寫作與總結的 AI 工具經常會改變文本的原意，以符合特定的政治觀點。例如，Meta、Google、阿里巴巴與 Mistral 的模型在氣候變遷與女性主義等議題上，傾向顯示自由主義觀點。相反地，xAI 的 Grok 則顯示出右翼視角，尤其是在墮胎問題上。即使 AI 被要求保留原意，這些改變依然發生，這表明這些模型具有內建的政治傾向。

Furthermore, researchers emphasized that these small changes could influence millions of users and, consequently, cause long-term changes in public opinion. They argued that this is a serious problem because current laws, such as the EU AI Act, are not strong enough to stop this. This means AI is no longer just filtering information but is actively rewriting it, acting as a gatekeeper of knowledge.

此外，研究人員強調這些細微的改變可能會影響數百萬名用戶，進而導致公眾輿論的長期改變。他們認為這是一個嚴重問題，因為目前的法律（例如歐盟 AI 法案）強度不足以阻止這種情況。這意味著 AI 不再僅僅是過濾資訊，而是在主動重寫資訊，扮演著知識守門人的角色。

At the same time, there is a huge increase in AI-generated content. Because LLMs are easy to use, people are producing massive amounts of spam. To fight this, platforms are using AI-driven detection systems. For instance, Reddit reported that its new AI tools reduced user exposure to spam by 20% between January and March. While YouTube and TikTok now require users to label AI content, experts assert that automated moderation still needs human supervision to be truly effective.

與此同時，AI 生成的內容大幅增加。由於 LLM 易於使用，人們產出了海量的垃圾訊息。為了對抗這一點，平台正使用 AI 驅動的偵測系統。例如，Reddit 報告其新 AI 工具在 1 月至 3 月期間，將用戶接觸到垃圾訊息的比例降低了 20%。雖然 YouTube 與 TikTok 現在要求用戶標記 AI 內容，但專家主張自動化審核仍需要人類監督才能真正有效。

The digital world currently faces two main challenges: the way AI introduces political bias into our conversations and the ongoing struggle between AI-generated spam and the tools used to detect it.

數位世界目前面臨兩個主要挑戰：AI 如何將政治偏見引入我們的對話，以及 AI 生成垃圾訊息與偵測工具之間持續的鬥爭。