OpenAI Announces Fix to ChatGPT’s Em-Dash Habit

New Delhi: OpenAI is bringing a solution to a problem that existed but was rarely spoken about in public - the inclusion of too many em dashes in ChatGPT’s writing. What seemed like a minor stylistic quirk has now been acknowledged by the company as a clear marker of AI-generated text today. 

Many ChatGPT users have pointed out that the repeated use of em dashes has become a giveaway. Readers could often identify ChatGPT’s output simply by noticing the punctuation pattern. 

OpenAI has now introduced a fix through its custom instructions feature, allowing users to tell ChatGPT not to use em dashes. The model adapts quickly when given this directive, producing text that feels more natural and less formulaic. 

This change is part of a broader push toward personalisation. By giving users control over tone, style, and formatting, OpenAI hopes to make ChatGPT more flexible and trustworthy. 

Reactions to OpenAI’s announcement about ChatGPT’s em-dash habit have been wide-ranging, with many users offering theories on why the punctuation became so common in AI writing.

Some believe the answer lies in the training data. “My guess is that books make up a huge portion of its training data, and since em dashes are common in books, it started adopting them,” one user wrote. Another suggested that the style emerged before Reddit posts were incorporated into training, noting that by the time those posts were added, they were already “AI slop.”

Others argue that em dashes were simply encouraged during reinforcement learning. “It’s just good writing that was encouraged during training and RLHF,” said one user, adding that they were “salty” about losing the option because em dashes made text flow better.

Some framed the issue in broader terms. One commenter described it as part of a corporate policy tied to long-term AI memory and industry-wide standards, while another said the punctuation was favoured because it made conversations sound more human and natural.

Technical explanations also surfaced. A few users pointed out that em dashes and similar symbols are more token-efficient, meaning the model could process them faster while keeping sentences connected. “Efficiency mostly — it’s the best way to keep a sentence or thought train flowing,” one user explained.

Read More: OpenAI CEO Sam Altman Bets Big on India, Says It's Going to Be ‘One of Our Biggest Partners’