ChatGPT will now respect “no em‑dashes” custom instructions — a small but revealing update

OpenAI CEO Sam Altman announced that ChatGPT will finally follow custom instructions that ask the model not to use em dashes. Previously, users reported that the assistant ignored this preference and continued inserting em dashes frequently. Altman called the change a “small‑but‑happy win,” reflecting a tweak to how the model honors formatting guidance.

Why does this matter? Many people have come to treat heavy em‑dash use as a quick heuristic for AI‑generated prose. While em dashes alone aren’t definitive proof of machine writing, their recurrent appearance in generative outputs made the punctuation a common “tell” among observers and journalists.

Why LLMs might overuse em dashes

• Training data: Large models learn from diverse sources — books, articles and forums — where em dashes often appear. Patterns in that data can lead the model to favor certain punctuation.
• Default style priors: Models implicitly adopt stylistic habits present in training text. Without explicit constraints, they may reproduce punctuation tendencies that look “AI‑ish.”
• Instruction handling: Earlier versions sometimes prioritized semantic instructions over minor formatting preferences, causing users to see ignored style requests.

OpenAI’s update suggests better compliance with user formatting preferences in custom instructions. However, punctuation‑based detection remains fragile: writers and editors should avoid assuming authorship from a single stylistic cue.

Practical takeaways

• If you dislike em dashes, add that preference to ChatGPT’s custom instructions — the model should now comply.
• Using punctuation or other surface cues to flag AI content can be helpful, but combine them with factual checks and provenance when accuracy matters.
• Writers can adopt deliberate style guidelines (e.g., prefer hyphens or parentheses) and include them as instructions to maintain consistent voice across drafts.

As generative systems improve, expect more granular control over tone and style in custom settings. Still, no single hack will reliably prove authorship — robust verification requires multiple signals and, where possible, metadata or provenance tools.

Discussion: Have you used punctuation or other style clues to detect AI writing? Will you ask ChatGPT for stricter style rules now that it appears to follow them? Share your experiences and tips.