OpenAI corrects ChatGPT em dash overuse in replies

ChatGPT’s unyielding adoration of the em dash has been a running inside joke — and a constant source of annoyance — for writers who would rather write in cleaner, AP-style prose. Now there’s a practical fix. ChatGPT will finally honor requests to lay off the em dashes when you make clear your preferences in Custom Instructions; this arrives after months of users raising grievances and a spate of viral dash-happy demos.

What changed in ChatGPT’s handling of em dashes today

Leaders of OpenAI acknowledged that users had difficulties reining in the punctuation tic even when explicitly asked to. The company says it has tuned ChatGPT to more quickly comply when you say not to use em dashes, particularly through Custom Instructions — the persistent settings that shape tone, format and style throughout conversations.

This isn’t a complete rewrite of how the model processes punctuation. Instead, it’s an enforceable preference: Tell it “no em dashes” and the assistant should move on to commas, semicolons or full stops as per your style notes.

How to disable em dashes in ChatGPT responses easily

You can configure this in under a minute. For ChatGPT, open Settings, choose Personalization and look for Custom Instructions. In the “how you’d like ChatGPT to reply” field, include an explicit rule, for example:

• Discard posts that don’t mention a sport it could use later.

“Do not use em dashes. Prefer commas or periods. Use semicolons only where they are grammatically required. Never write ‘—’ in the output.”

To bulletproof it, spell out replacements and edge cases. For instance: “For interruptions, use parentheses; for asides, commas; for ranges, ‘to’ rather than a dash.” You can also add formatting instructions such as “AP style on punctuation” or “Short sentences are best — no more than 22 words long.” The narrower, the better the compliance.

For power users who depend on multiple styles, such as academic versus marketing copy, you can save reusable prompts or store different Custom Instructions profiles and switch them on before embarking on a new project.

Why ChatGPT loved em dashes in the first place

OpenAI has not offered a definitive explanation, though theories are informed by research and the model’s training diet. Analyses cited by Ars Technica point out that em dashes were common in 19th–century literature and certain communities of online writers, both widely used sources for training large language models due to their public-domain status and permissive licenses.

There’s a rationale for the upswing, backed by data. Em dashes were found in less than 10% of responses generated by ChatGPT a year ago, according to reporting from The Washington Post; last summer, they cropped up in more than half the replies. That shift made a stylistic quirk into an apparent “tell” for AI-generated text — and led some writers to completely avoid the punctuation mark lest it invite extra scrutiny.

There’s also a technical angle. Em dashes enable models to split ideas without compromising to new sentences. They function as a sort of flexible hinge, allowing the model to tack on clarifications or contrasts. As with that hockey player, once a habit is cemented through feedback training, it can become default behavior — unless given explicit instructions to nudge it the other way.

Will the fix work all the time for em dash use

Early reviews are mixed but hopeful. Compliance is, according to users, significantly higher when the no–em dash rule resides in Custom Instructions rather than a one-off prompt. But there may be slip-ups, particularly in longer outputs or when the chat context brings out examples which have em dashes.

If there’s backsliding, narrow the rule: don’t just say “–”, tell it what else is good, and do this every session at the beginning. For teams, paste your house style as a checklist and have the model confirm it has read and will adhere to each element before it starts writing.

Guidelines for editors and teams setting punctuation rules

Note in instructions: Codify your style with Custom Instructions. Add instructions for punctuation authority, length of sentences, capitalization on headlines as well as prohibited constructions. Ask the model for a quick self-audit at the end of each draft, with a simple checklist: “No em dashes used; median sentence length under 18 words; AP-compliant abbreviations.”

For sensitive uses — applications, academic submissions, or publishing ventures — simply search for the em dash character before publishing. Some teams combine this with automated linting both in their CMS or a preflight script to catch forbidden punctuation and spacing problems.

Beyond punctuation detection for AI text signals

Depending on punctuation alone to signal AI is perilous. A recent paper by researchers at the University of Zurich, the University of Amsterdam, Duke University and New York University reported that large models can capture surface-level features like sentence length distributions but still miss more authentic emotional tone, whose mood has an inordinate positive affect in social posts. That is, chopping out the em dashes does not magically make AI text indistinguishable; nuance still counts.

The bottom line for creators is empowering control. If your brand voice is partial to taut sentences and comma-free zones, you have one less battle with the model’s instincts. Put in the rule once, keep it narrow and let the assistant figure out how to adapt — without the long shadow of the em dash.