ChatGPT Achieves a New Level of Intelligence: Not Using the Em Dash (gizmodo.com)

OpenAI quietly scored a “small-but-happy” win: GPT-5.1 can now follow a user’s custom instruction to avoid using the em dash. CEO Sam Altman announced the change after the company demonstrated the model obeying a “no em-dash” directive in personalization settings — even though an earlier apology example still needed an em dash to be written. This is not a universal rewrite of punctuation behavior; it’s a per-user control that biases the model’s output when that instruction is provided. Technically, the change appears to come from weighting custom instructions more heavily during response generation rather than fixing a core text-generation tendency. That aligns with OpenAI’s rollout message that GPT-5.1 improves instruction-following and personalization. The result highlights two broader points for the AI community: fine-grained, instance-level controls can mitigate annoying alignment quirks without retraining models, but variability across users shows the limits of that approach and the opaque, hard-to-debug nature of large language models. In short, it’s a useful UX improvement and a case study in personalization-first strategy — not evidence of progress toward AGI.