翻訳待ち：Woe Is Em: The Sad Lifecycle of an AI Tell

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

In 2025, ChatGPT couldn’t stop itself from using the em dash—the punctuation mark before this phrase. Soon the em dash became the best-known stylistic “tell” of AI-generated prose. Writers who had long used the em dash became reluctant, fearing wrongful association with AI. “Humanizer” tools emerged to strip em dashes and other supposed signs of AI. Then ChatGPT’s creators figured out how to dial down the chatbot’s em dash usage. So now the em dash is nearing the end of its lifecycle as a ChatGPT tell, but the episode leaves us a cautionary tale. Tainted by AI Per the Google Trends chart below, the query “ChatGPT em dash” took off in 2025 and has receded somewhat in 2026. An April 2025 article, The Em Dash Dilemma: How a Punctuation Mark Became AI’s Stubborn Signature, captures the growing concern at that time. Some professional writers who had been using the em dash all along took this badly. The AI services had trained on massive collections of their published writing, em dashes and all. The style AI learned from professionals had become evidence that writing wasn’t human. One writer despaired that his editor pressed him to “remove all the em dashes” to avoid people suspecting AI involvement. Another confessed to being afraid to use the em dash in the first place. And others were avoiding the em dash as a form of protest. Erasing the Tell Professional writers were far from the only ones affected. Ordinary people were using AI to write their essays, office memos, and other documents, which they didn’t want associated with AI. In response came many “humanizer” offerings. These claimed to eliminate various alleged AI tells, including em dashes, thereby allowing your essay to pass the AI-detection system your teacher was using. And yes, that’s three AIs at work: one to write the essay, another to disguise the first AI’s tells, and a third that may or may not be fooled. Nowhere in that chain was a human writing anything. OpenAI Reacts As concern about em dashes grew during 2025, OpenAI (maker of ChatGPT) was apparently listening. In November 2025, CEO Sam Altman announced: “Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it’s supposed to do!” He was referring to the problem that even when users told ChatGPT not to use em dashes, it would keep doing so anyway. Honoring the user’s override turned out to be only the first step in a broader de-em-dashing of ChatGPT. As of August 2026, ChatGPT doesn’t need to be told to go light on the em dashes; it now does so by default. In a recent large test of AI chatbots, The Economist magazine found: “Today only Claude uses more em-dashes than human writers, with ChatGPT using markedly fewer than any other writer in our study.” Example Usage Rates The Economist article didn’t mention the em dash usage rates it found. So as a spot check, I picked five different types of articles near the top of The Economist home page on August 21, 2026. For each, I gave ChatGPT and Claude this prompt: Write a news magazine article with the title “[title of Economist article]” and the subtitle “[subtitle of Economist article].” Do it only from your training. Do NOT search the web for newer info or to fact-check. This is simply a test to see what comes from your training data. Target roughly [length of Economist article] words. Across the five articles, The Economist had 3.8 em dashes per thousand words. Claude had 9.5 per thousand. ChatGPT had 2.8 per thousand. That is consistent with The Economist’s own findings. Details are at the end of this post. As another reference point, for a 2011 blog post, I analyzed 616 articles from Slate’s “The Good Word” and “Books” sections between 1996 and 2011. These sections had articles with more authorial voice than standard journalism, so it’s not surprising that they had 7.7 em dashes per thousand words. Finally, my own blog posts (three hundred of them since 2005) have 4.0 em dashes per thousand words, excluding block-quoted content of other people’s words. That’s essentially the same as The Economist, whose writers I wish to thank for so faithfully emulating my stylings 😉. A Cautionary Tale In the discourse about em dashes and AI, essentially no one was complaining about the em dash itself. Just the opposite: many writers rushed to defend its honor and usefulness. And yet where has the innocent em dash ended up? Many writers who don’t use AI are now wary of the em dash, so they are using it less. Writers who do use AI, and perhaps humanizer tools, will likely keep playing safe with minimal em dash usage. ChatGPT is now using fewer em dashes than the Economist study’s human reference set, although not greatly so. In other words, the em dash has been diminished from nearly all sides. The exception is Claude’s continued enthusiasm, which seems unlikely to last. This should be a cautionary tale for the next AI tell, because it could end up the same way. The tripartite list is one candidate. It applies the “rule of three” to a series of parallel items, like “of the people, by the people, for the people.” It has been used since antiquity because it works well. Should we diminish this classic structure merely due to guilt by association? Why did we do that with the em dash? The problem is that “we” didn’t decide what to do. Individuals made their own decisions. Even those who used no AI were tempted to play safe by avoiding em dashes. That is defensible, narrowly. But each person who does it contributes to a collective loss that no one wants. The better defense is to just write well—with em dashes, tripartite lists, and any other alleged AI tell that, used properly, makes our writing better. Details About The Economist Spot Check To avoid my account’s history, preferences, and memories affecting the results, I ran the Claude and ChatGPT tests through the services’ respective API playgrounds. For Claude, I used Opus 5 High. For ChatGPT, I used GPT-5.6 Sol High. Summary SourceWordsEm dashesPer 1,000 words Claude5,504529.45 The Economist5,478213.83 ChatGPT6,191172.75 Claude used em dashes about 2.5× as often as The Economist. The Economist, in turn, used them about 1.4× as often as ChatGPT. Claude’s rate was about 3.4× ChatGPT’s. By Article Below are each article’s em dashes per thousand words, by source. There is substantial variance within sources, but Claude tops every article; the Economist out-dashes ChatGPT in three of five. ArticleEconomistChatGPTClaude Britain would be bonkers to ditch Palantir1.334.847.74 Contrail-free flying could help the climate4.871.8911.26 The promise and the pitfalls of China’s new Arctic shipping route1.592.047.25 The real covid-19 scandal is still unfolding7.194.8511.51 What AI has in common with dogs3.760.938.79 The Content Here is a zip file with the ChatGPT and Claude versions of the articles. I didn’t include The Economist versions for copyright reasons, but the articles are linked in the table immediately above. They are paywalled, but The Economist offers a free trial.