OpenAI says it has finally tamed one of ChatGPT’s most famous stylistic tics: the em dash runneth over.
CEO Sam Altman added that the model now consistently respects requests to avoid em dashes when users indicate this in Custom Instructions, which he noted is a small but nice quality-of-life improvement.
It’s a surprisingly consequential fix. The em dash had become something of a pop-culture tell for AI-written text, showing up in emails, student papers, marketing copy and support chats. Although many a human writer has felt the love for the mark, this refusal of the chatbot to avoid it — even when asked — supplied suspicion and false alarms.
What Changed and Why It Matters for Users
OpenAI’s update focuses on steerability. If you tell ChatGPT in Custom Instructions to “avoid em dashes” or that “you should use commas or semicolons instead,” it now does so much more reliably. The company also conceded in Threads that the model had made the em dash uncool through sheer omnipresence and “apologized,” making it clear that user preferences would be honored from now on.
Indeed, this means writing style is not left up to a coin flip. For teams drafting those customer communications or for educators grading papers, fewer “accidental em dashes” mean less confusion and the appearance of A.I. fingerprints. It’s not like a default ban — you can still use em dashes if that’s what you want, but when you go to turn the knob, it finally begins to move.
Why Em Dashes Became a Common A.I. Tell in Writing
Large language models are trained on large text corpora, which use the em dash frequently, particularly in essays, blogs and long-form commentary. In training — reinforcement learning — the model tends to prefer punctuation that neatly sews clauses together in a way that is not quite like commas, colons or parentheses. The em dash does that beautifully — and quickly — so it was a safe, high-probability guess.
The catch: readers inferred em dashes as a stand-in for AI. That’s a blunt instrument. Style guides such as The Chicago Manual of Style and the AP Stylebook approve their use in moderation, and many journalists and authors rely on them for rhythm. Research communities such as Stanford HAI and The Allen Institute for AI have cautioned against them since superficial cues are not reliable signals of machine authorship. Likewise, education vendors like Turnitin report false positives returned by detection tools, especially for English as a foreign language writers — as if we needed another reason not to trust single-cue judgments.
How to Nudge ChatGPT Away From Em Dashes
Preferences can be set in the personalization settings of ChatGPT under Custom Instructions. In general, plain-language prompts work well: “Please avoid em dashes and opt for commas or periods,” or “No em dashes; use parentheses judiciously.” Teams: Standardizing a shared “instruction snippet” like above helps keep brand voice consistent across drafts.
A quick one: instead of “We’ve shipped the update–thanks for your patience–we’ll monitor performance,” ask “We shipped the update. Thanks for your patience. We’ll monitor performance.” The model shall now obey that without going against it in the middle of a paragraph.
AI Detection and the Broader Trust Implications
Detectors and ad hoc classroom heuristics that were somehow leaning on punctuation patterns have just lost another piece of crutch. And as the models turn more and more steerable, these surface-level tells will (presumably) wane. That’s a healthy thing: Slightly more rigorous verification should be deploying provenance tools, process transparency and content vetting — not depending on whether a sentence here prefers a dash to a comma there.
For communicators, it also minimizes the need to “edit out” AI-isms that make their way into drafts. Editors can concentrate on substance, sourcing and evenness rather than mere mechanical search-and-replace. And for writers who are fond of the em dash? Nothing changes; you can still have it in your toolkit, and ask the model to reflect that style.
The Bigger Picture for Style Control in AI Writing
The em dash fix is a microcosm of that larger frontier: controllable generation. Enterprises want models that conform to tone guides, legal guidelines, and accessibility standards when you need them. OpenAI’s transition hints at continued investment in following fine-grained instructions — right down to the frequency of punctuation marks. Anticipate similar control with hedging phrases, sentence length — maybe even regional spelling.
No model is perfectly obedient, and decoding settings can always push toward certain phrasing. But in that sense, this update moves ChatGPT closer to a reliable writing partner — one that puts down the words you tell it to write, not the ones internet jokers taught it how to say.
