OpenAI has addressed a longstanding writing quirk in ChatGPT: the overuse of em dashes. CEO Sam Altman tweeted that users can now instruct ChatGPT not to use em dashes through custom instructions. “Small-but-happy win,” he wrote, noting the update finally works as intended.
On Threads, the @ChatGPT account added that users can access this feature via Settings > Personalization > Custom Instructions. The change comes after writers increasingly flagged em dashes as a potential AI “tell” in their work, with many avoiding them to prevent misattribution.
Em dashes appeared in less than 10% of ChatGPT responses a year ago but over half of responses by summer 2025 included them, according to The Washington Post. OpenAI hasn’t explained why, but experts suggest it may stem from the prevalence of em dashes in 19th-century books and popular Medium blogs, which are common sources of training data.
While the fix allows for alternative punctuation like short hyphens, researchers caution that AI detection remains complex. A recent study from the University of Zurich, Amsterdam, Duke, and NYU noted that LLMs can mimic structural features like sentence length but still struggle to produce believable emotional tone on social media platforms.
