WRITING IS HARD. For Franz Kafka, putting words on the page was a process of “unending torments”. George Orwell said it felt like “a long bout of some painful illness”. Hunter S. Thompson, a journalist, saw it as the opposite of sex: “only good when it’s over”. Now AI has come to the rescue. Large language models (LLMs) eagerly offer reams(大量（纸张或文字）) of text without any pain or torment. Bots can spew(大量喷出（常含贬义）) out emails (excellently), longer prose (passably) and poetry (poorly). To them, writer’s block is a 500-sheet pack of paper: any user of an LLM can produce as many words in a day as Kafka did in his life.(任何LLM用户一天就能写出卡夫卡一生所写的文字。)
写作是艰难的。对弗朗茨·卡夫卡而言，将文字付诸纸面是一个“无尽折磨”的过程。乔治·奥威尔说这感觉像“一场漫长而痛苦的疾病”。记者亨特·S·汤普森则认为写作是性的反面：“只有在结束时才是好的”。如今，人工智能前来救场。大型语言模型（LLM）热切地提供大量文本，毫无痛苦或折磨。机器人可以吐出电子邮件（出色）、较长的散文（尚可）和诗歌（糟糕）。对它们来说，写作障碍就是一包500张的纸：任何一个LLM用户一天内就能写出卡夫卡一生所写的那么多文字。
For human writers this is a blessing and a curse(既是福也是祸). If you claim to be an author but use AI, you look like a fraud(骗子；欺诈行为)—bot-busting(打击机器人的) tools like Pangram(一款AI检测工具) are good at spotting when someone has had an AI helping hand(帮助；援手). Take the hundreds of people who have used LLMs to write applications for The Economist’s internships, or the government ministers who have submitted AI-generated works to our By Invitation column(《经济学人》的“特邀”专栏) (it’s tempting, but we won’t name them).
对人类作家来说，这既是福也是祸。如果你自称是作者却使用AI，你会显得像个骗子——像Pangram这样的反机器人工具很擅长发现有人借助了AI之手。以数百名使用LLM撰写《经济学人》实习申请的人为例，或者那些向我们的“特邀”专栏提交AI生成作品的政府部长们（虽然很想点名，但我们不会这么做）。
Our investigation of the stylistic quirks(怪癖；独特之处) of the big bots reveals how, with every update, AI writing is getting closer to human writing. We compared the outputs of LLMs to prose we know is produced by humans: our own.(我们将LLM的输出与人类创作的散文——我们自己的作品——进行了比较。) We asked 14 variants of ChatGPT, Claude, Gemini and Grok released since 2024 to write versions of our articles without consulting the web. We compared our corpus(语料库) across 55,940 sentences and 1.2m words, and checked our results against other newspapers and hit fiction books(畅销小说). We found that some of AI’s former tricks—using lots of em-dashes(长破折号), for instance, or words like “leveraging(利用（杠杆作用）)”—are no longer leveraged much in its writing.
我们对各大机器人风格怪癖的调查揭示了，随着每一次更新，AI写作正越来越接近人类写作。我们将LLM的输出与我们知道是人类创作的散文——我们自己的作品——进行了比较。我们让自2024年以来发布的14个版本的ChatGPT、Claude、Gemini和Grok在不查阅网络的情况下撰写我们文章的版本。我们比较了涵盖55,940个句子和120万词的语料库，并将结果与其他报纸和畅销小说进行了核对。我们发现，AI以前的一些惯用伎俩——例如大量使用破折号，或像“leveraging”这样的词——在其写作中已不再那么频繁使用。