人工智能大语言模型（LLM）的兴起极大降低了内容创作的门槛，使用户能够瞬间生成海量文本。然而，这种便捷性也带来了严重的垃圾内容泛滥风险。根据最近的一项研究，LLM使亚马逊每月发布的新电子书数量翻了三倍，达到30万本，并且在一个统计中起草了超过三分之一的新网站。为了评估AI写作风格的演变，《经济学人》通过分析自2024年以来发布的包括ChatGPT、Claude、Gemini和Grok在内的14个模型版本，对比了55,940个句子和120万个单词，发现AI正在逐步摆脱过往过度使用长破折号或“leveraging”等特定词汇的痕迹，写作质量正日益逼近人类专业水平。
尽管AI模型生成的文本日益自然，但鉴别工具（如Pangram）依然能够精准识别其痕迹。在《经济学人》收到的实习申请和政府官员提交的专栏文章中，频繁出现由AI生成的痕迹。AI写作的主要缺陷在于过于冗长、过度依赖拉丁语词汇、滥用陈词滥调以及频繁使用“三分法”修辞等。例如，AI常生成如“该谜团已被解决——不是部分地，不是暧昧地，而是明确地”等冗余句子。这种膨胀而机械的文本不仅给读者带来认知负担，也使得人类在内容生产流程中的角色从撰写者加速转向编辑者。
为了有效地将AI生成的初稿转化为高质量的文本，人类编辑需要掌握三大核心策略。首先是精准下达指令，通过清晰的要求和反复的迭代提示来指导模型输出。其次是注重细节核查，精简机械重复的修辞句式，替换晦涩的术语并剔除虚浮的陈词滥调。最后是准确把握受众需求，矫正AI过于谄媚或好为人师的语气，例如在撰写严肃邮件时要求其取消感叹号。随着AI写作能力的持续提升，人类掌握精细的编辑技巧将成为保持内容真实性与高品质的关键所在。
The rise of Large Language Models (LLMs) has fundamentally altered the mechanics of text generation, enabling users to produce vast amounts of written content instantaneously without traditional creative block. However, this ease of generation has unleashed a flood of automated material across digital media. According to a recent study, LLMs have tripled the volume of new e-books published on Amazon each month to 300,000 and now draft more than one-third of all new websites. To track the evolution of automated writing, an investigation analyzed 14 model variants of ChatGPT, Claude, Gemini, and Grok released since 2024 across a corpus of 55,940 sentences and 1.2m words, finding that AI writing is rapidly shedding earlier stylistic markers like excessive em-dashes and jargon like "leveraging" to closely mimic human prose.
Despite these stylistic advancements, automated text continues to display distinct shortcomings that require human oversight. Advanced detection tools like Pangram easily spot AI contributions in submitted materials, including hundreds of internship applications and minister-level guest essays sent to major publications. AI-generated prose frequently suffers from verbosity, heavy reliance on Latinate terminology, clichéd metaphors, and an over-application of the "rule of three" rhetorical device. For instance, models routinely produce redundant phrases such as claiming a mystery is solved "not partially, not ambiguously, but definitively." Such structural flaws add unnecessary cognitive friction for readers, elevating the importance of human editorial intervention.
To effectively transform raw machine output into compelling prose, human editors must employ three key strategies. First, editors should commission thoroughly by providing precise instructions and iterating prompts repeatedly until the machine refines its output. Second, editors must pay strict attention to detail, removing Latinate jargon, eliminating redundant rhetorical triplets, and replacing empty buzzwords with direct language. Finally, editors must tailor the output to the intended audience by stripping away the sycophantic, overly enthusiastic, or preachy tone that LLMs default to. As AI text generation becomes increasingly ubiquitous, sharp editorial judgment remains the ultimate filter for quality and authenticity.
Source: AI is getting better at writing. Humans must get better at editing
Subtitle: Our investigation suggests how
Dateline: 7月 30, 2026 05:37 上午