AI often makes writing worse. (Even if it makes it better) Why is complicated. But it’s not about the damn em-dash Sep 24th 2026|4 min read Writing using artificial intelligence is everywhere. It has crept into elite newspapers (the* Wall Street Journal* ran an op-ed drafted by it) and elite publishing houses (Grasset in France finds itself defending a star author accused—falsely, he claims—of using it in a prize-winning novel). It is in the speeches of British politicians, in the homework of British students and abundantly there in LinkedIn posts (41% of longer ones, according to a study by Pangram, a detector, use it). Ask Gmail to draft a love letter to your partner and it will instantly offer up “My Dearest [Husband’s Name]…”. It is quick. It is not quite Keats. Who cares? The reactions show that people do care. AI-spotting has become a mania. Prose is being scrutinised with the attention, and jargon, once reserved for Latin literature. Arcane rhetorical terms like antithesis and epanorthosis (the “not X, but Y” construction) and tricolon (threefold things, like “veni, vidi, vici”) are bandied about by social-media users with the verve of Ciceronian scholars. Literary criticism has rarely been so hot. The em-dash, the most notorious AI “tell”, seems to be acting as—another rhetorical word—a synecdoche, the term for when writers use a part for the whole, like “wheels” for “car” or, perhaps, “em-dash” for “the intellectual obsolescence of the human race”. This, says Maryanne Wolf, an emeritus professor of neuroscience at UCLA, is not a battle over punctuation. It is a “battle for our minds”. AI is so “unprecedented”, says Steven Pinker, a professor at Harvard, that we cannot tell how, or whether, it will affect the English language. Happily there is a body of evidence dating back a few years (about 1,300) on what good writing is. It is called “English literature”. The first thing that a millennium of English writing shows is that most people are terrible at it. They find it painful to do (“like a long bout of some painful illness”, as George Orwell had it) and produce words that are almost as painful to read. AI can be bad; humans are often worse. Dan Brown offered: “His eyes went white, like a shark about to attack.” William Wordsworth offered an ode to a spade (it began “Spade!”). Lee Child offered the phrase: “It was about as distinctive as the most distinctive thing you could ever think of.” AI, by contrast, writes with the ease of the easiest thing you could ever think of. It can write a sonnet in half a second, a 1,000-word article in ten seconds, an Austen-length novel in half an hour. And though it can be bad, it often isn’t. In a recent** blind tasting** of two short stories, one human-written and one AI-written, researchers from Villanova University found that respondents rated the AI story as “higher quality and more absorbing”. Yet English literature also shows that writing is not just the act of spewing words onto a page. Writing, as writers point out, is no more merely words on the page than a smile is joy, or a tear is sorrow. Real writing is not writing at all. It is thinking; specifically it is thinking: “What do I think?” Often “you only really know what you think by writing it,” says Emma Smith, a professor of English at Oxford. For how, as a character quoted by E.M. Forster puts it, “can I tell what I think till I see what I say?” When Michael Gove, the editor of the Spectator and an ex-MP, was in the cabinet he told his underlings to “Think in ink.” To make a good argument you need to have followed its logic “step by step”—in writing. Those who use AI to write speeches are, he says, using “counterfeit currency”. But counterfeit currency is always disliked; AI writing is not. Few would object to someone using AI to contest a parking fine, fill in a tax return or for a second-language speaker to polish their English. If a company uses AI to announce a new parking policy, few care. But, says Colleen Kirk, a professor at the New York Institute of Technology, if that policy was problematic and the company apologised using AI—then “You’ve got a big problem.” You risk activating “moral disgust”. In 1970 a Japanese robotics expert, Mori Masahiro, noted that the more human a robot looks, the more** affinity **humans feel for it—until it becomes too human, whereupon affinity turns to revulsion: the *“uncanny valley”. If you shook someone’s hand in a darkened room and, unexpectedly, clasped a prosthetic hand you would, Mori wrote, “shriek*”. AI writing—bland, plasticky, human-ish—seems to sit in the “uncanny valley” of prose. It certainly makes social media shriek. So here is the AI paradox. AI prose is easier to produce and often better than human prose. Yet humans do not like it; they may** impair **their ability by using it and they often loathe its use by others. This may be ridiculous but is encouraging. As humanity destroys its intellectual dominion with its own intellect—as it is beaten back in chess, in Go, in coding—we retain one last advantage: we are human. And, for now, humans (often) prefer to read humans. So make the most of that and, as E.M. Forster also said, *“Only connect”*. And probably write personal emails yourself. Especially to your “[Dearest Husband]”. ■ 使用人工智慧來寫作無所不在。它已經悄悄溜進了頂尖報紙（《華爾街日報》刊登了一篇由它起草的專欄文章），也滲透進了頂尖出版社（法國的格拉塞出版社發現自己不得不為一位明星作家辯護，該作家被指控——但他聲稱這是誣陷——在一本獲獎小說中使用了 AI）。它出現在英國政治人物的演講稿中、英國學生的作業裡，更大量充斥於 LinkedIn 的貼文中（根據檢測工具 Pangram 的研究，較長的貼文當中有 41% 使用了 AI）。如果你要求 Gmail 給你的伴侶寫一封情書，它會立刻為你呈現出：「我最親愛的［丈夫姓名］……」。它非常快速。但它並不太像濟慈（Keats）(英國著名詩人)的詩句。 誰會在意呢？人們的反應顯示出，大家其實非常在意。尋找 AI 的痕跡（AI-spotting）已經變成了一種狂熱。人們正在用過去專門留給拉丁文學的專注力與專業術語，來嚴格審視這些散文內容。諸如「對照法」（antithesis）、「更正修正法」（epanorthosis，即『不是 X，而是 Y』的句型結構）以及「排比法」（tricolon，即三重疊句，例如『我來，我見，我征服』）等晦澀的修辭術語，被社群媒體用戶以宛如西塞羅學者般的熱情隨口引用 (西塞羅--馬庫斯·圖利烏斯·西塞羅（Marcus Tullius Cicero，西元前106年－前43年）是古羅馬共和國晚期的哲學家、政治家、律師、作家與雄辯家。他奠定古典拉丁文學風格)。文學批評很少像現在這樣紅火過。 破折號（em-dash）作為最惡名昭彰的 AI「辨識特徵」（tell），似乎正充當著——這裡又是另一個修辭學詞彙——「提喻法」（synecdoche），也就是創作者用局部來代表整體的用語，比如用「輪子」代表「汽車」，或者也許是用「破折號」代表「人類智力的過時」。加州大學洛杉磯分校（UCLA）的神經科學榮譽教授瑪莉安·沃爾夫（Maryanne Wolf）表示，這並不是一場關於標點符號的爭奪戰。這是一場「爭奪我們心智的戰鬥」。 哈佛大學教授史蒂芬·平克（Steven Pinker）表示，AI 是如此「前所未有」，以至於我們無法預測它將如何影響、或是否會影響英語這門語言。幸運的是，關於什麼是好文章，我們擁有追溯至好幾年前（大約 1,300 年）的大量證據。那被稱為「英國文學」。 上千年來的英語寫作所揭示的第一件事，就是大多數人其實都極度不擅長寫作。他們覺得寫作是一件痛苦的事（正如喬治·歐威爾所說的，『就像經歷了一場某種痛苦重病的長期發作』），而且寫出來的文字讀起來也同樣令人痛苦。AI 可以寫得很爛；但人類往往更爛。丹·布朗（Dan Brown）曾寫過這樣的句子：「他的雙眼翻白，就像一條即將發動攻擊的鯊魚。」威廉·華茲華斯（William Wordsworth）寫過一首獻給鏟子的頌歌（開頭竟是『鏟子啊！』）。李·查德（Lee Child）則寫出了這樣的短語：「這大概跟你所能想到的最特別的事物一樣特別。」 相比之下，AI 寫起文章來，簡直就像你能想到的最輕鬆的事物一樣輕鬆自如。它可以在半秒鐘內寫出一首十四行詩，在十秒鐘內寫出一篇 1,000 字的文章，在半小時內寫出一本珍·奧斯汀體量的小說。而且雖然它可能會寫得很爛，但通常並不會。在維拉諾瓦大學（Villanova University）的研究人員最近進行的一項針對兩篇短篇小說的盲測中（一篇由人類撰寫，一篇由 AI 撰寫），受訪者評定 AI 寫的小說具有「更高的品質且更吸引人」。 然而，英國文學也表明了，寫作不僅僅是把文字噴吐到頁面上的行為。正如作家們所指出的，寫作絕不僅僅是頁面上的文字，就像微笑不單純等同於喜悅、眼淚不單純等同於悲傷一樣。真正的寫作根本不是寫作本身。它是思考；具體來說，它是去思考：「我到底具體是怎麼想的？」 牛津大學的英語教授艾瑪·史密斯（Emma Smith）說，通常「你只有透過把想法寫下來，才能真正知道自己到底在想什麼。」因為正如 E.M. 佛斯特（E.M. Forster）作品中的一個角色所說的：「在我看到自己說了什麼之前，我怎麼知道自己是怎麼想的？」當《旁觀者》（Spectator）雜誌編輯兼前國會議員麥克·戈夫（Michael Gove）還在內閣時，他曾告訴他的下屬要「用墨水來思考」（Think in ink）。要提出一個好的論點，你必須在寫作中「一步一步地」遵循其邏輯。他說，那些使用 AI 來撰寫演講稿的人，使用的是「偽鈔」。 但偽鈔總是令人厭惡的；AI 寫作卻並非總是如此。很少有人會反對有人利用 AI 來申訴停車罰單、填寫報稅單，或是讓非母語人士用它來潤飾英文。如果一家公司使用 AI 來宣布一項新的停車政策，很少人在乎。但是，紐約理工學院的教授柯琳·柯克（Colleen Kirk）表示，如果這項政策出了問題，而公司使用 AI 來道歉——那麼「你就遇到大麻煩了。」你有可能會觸發人們的「道德反感」。 1970 年，日本機器人專家森政弘注意到，機器人看起來越像人類，人類對它感到的親和力就越強——直到它變得「太像」人類，親和力便會驟然轉變為反感：這就是「恐怖谷」（uncanny valley）。AI 的寫作——平淡、帶有塑膠感、似人非人——似乎正處於散文風格的「恐怖谷」之中。它確實讓社群媒體尖叫不已。 所以這就是 AI 的悖論。AI 的散文更容易產出，而且往往比人類寫的散文更好。然而，人類並不喜歡它；人們可能會因為使用它而削弱自身的思考能力，而且他們常常厭惡其他人使用它。這或許很荒謬，但卻振奮人心的。當人類用自己的智慧摧毀自己的智慧主導權時——當我們在西洋棋、圍棋、程式碼編寫上被擊敗時——我們依然保留著最後一項優勢：我們是人類。而且，至少在目前，人類（通常）更喜歡閱讀人類寫的東西。所以，好好利用這一點吧，正如 E.M. 佛斯特也曾說過的：「唯有連結」（Only connect）。而且，大概還是親自撰寫你的個人郵件吧。特別是寫給你的「［最親愛的丈夫］」。■ --Hosting provided by SoundOn