OpenAI hasn’t yet achieved its goal of developing artificial general intelligence (AGI) or an autonomous “AI researcher.” However, the company has found a way to prevent ChatGPT from misusing the “em dash.” CEO Sam Altman announced this development in a post on the X platform. Altman said, “If you tell ChatGPT not to use em dashes in your specific instructions, it finally does what it’s supposed to!” He called this development a “small but happy win.”

ChatGPT’s annoying obsession with em dashes finally has a solution

The company confirmed this new capability in a post on Threads. In this post, they formally apologized to ChatGPT for “ruining the em dash.” Interestingly, even while writing this apology, ChatGPT couldn’t resist using em dashes. This reveals an important distinction about the solution.

In fact, OpenAI hasn’t made ChatGPT use em dashes more appropriately or less favorably by default. Instead, it has given users the ability to tell them not to use them through the chatbot’s customization settings. This means it’s more of a user-specific adjustment option than a solution to a fundamental problem with the model.

This new capability follows the release of OpenAI’s latest model, GPT-5.1. One of the key improvements the company emphasized in its introduction of the new model was that GPT-5.1 is better at following instructions and offers more customization features. The em dashes restriction is more of an example of how users can benefit from this “more adaptable” structure of the model, rather than a fix to the model’s underlying output.

The fact that this adjustment must be made on a user-by-user basis highlights how “black boxes” most large-format language models (LLMs) still remain. In fact, some users responding to Sam Altman’s post on X pointed out that their version of ChatGPT still uses em dashes despite specific instructions. OpenAI’s presentation of the solution as personalization suggests that finding a scalable solution remains challenging.

The company appears to have found a way to prioritize specific instructions in its calculations when generating a response. However, they still don’t seem to have fully grasped why the problem arose in the first place, or why it persists. Perhaps this is why the company has been less vocal about Artificial General Intelligence (AGI) lately and more focused on personalization features.

ChatGPT’s persistent use of em dashes can be particularly frustrating for content creators. Have you noticed this issue while using ChatGPT? Do you find the company’s personalization solution sufficient?