Language has habits.
Each person has a different rhythm, and even when using the same expression, there is a subtly different texture that lingers. That writing remains a human act is ultimately because of these minute traces.
And now, AI is erasing those traces one by one. OpenAI's recently announced improvement to em-dash control looks trivial at first glance. The punctuation mark that kept appearing even when users prohibited it — the em-dash.
In English-speaking communities and educational settings, this mark had been treated like a "fingerprint of AI-written text." Because it kept appearing even when prohibited. So when the em-dash appeared in students' submitted assignments, it was like receiving a guilty verdict that AI had written it.
This way, after the generative AI revolution, a single punctuation mark — the em-dash (—), commonly called a hyphen — created a bizarre scene where it stood at the center of debates over technology, education, censorship, and transparency.
Then on the afternoon of November 14, 2025, Sam Altman left a single sentence on X.
"Small but happy victory. Now if you ban it, it actually doesn't use it."
It was just one line, but the meaning was not small. Because OpenAI was declaring it had finally succeeded in controlling the model's stylistic instinct.
Asking AI to write something is no longer unfamiliar.
We entrust sentences to generative AI, ask it for ideas, have it summarize, and refine formats. This flow is an irreversible tide of the times.
I rather think such acts should become even more active. Because when we delegate to machines what machines can do, humans gain the leisure to ask deeper questions.
But hearing the news that the em-dash — that punctuation mark that kept appearing even when ChatGPT users prohibited it — can finally be completely controlled, I found myself thinking a lot.
Because the em-dash controversy is not a matter of simple convenience features but an event that revealed the structural properties of how AI writes.
For a time, the em-dash (—) was treated as the fingerprint of AI writing.
Students were suspected when em-dashes appeared in their assignments, and the marketing industry tried to detect AI usage just from message style. A single punctuation mark had become the criterion dividing technology and society, education.
But now that criterion has collapsed.
AI has reached the stage of erasing its own stylistic habits and controlling even individual punctuation marks to match user requirements. On the surface this looks like a trivial feature improvement, but it means that AI's style instinct — and the way it had been mimicking human traces — has become one step more refined.
Looking at the technical analysis, the significance of this change is even greater.
The em-dash issue is not a simple "punctuation bug." It is an event that revealed the structural properties of large language models.
First. The problem of LLM style habits.
In high-quality English-language writing materials, em-dashes are very common. Considering the proportion of academic papers, essays, and columns in the model's training data, it is a natural result that the model internalized this mark as its default style. In other words, the em-dash was a mark that ChatGPT believed was its own 'literary rhythm.'
Second. The problem of user instruction priority.
LLMs have a structure where they first consider rules for making sentences natural and then reflect user commands. So it appears that rules for making the flow of writing smooth had higher priority than the user's request of "don't use em-dashes."
Third. Subtle tension with AI detection technology.
Some detection tools have used em-dash usage as a pattern of AI text. Therefore, the em-dash suppression feature contains a structural challenge that conflicts with detection technology. A feature that erases AI's traces is simultaneously a feature that weakens the credibility of AI detection technology.
OpenAI's success in solving this challenge this time means they have sophisticatedly redesigned the style control layer inside the model. This is not a simple feature improvement but a technological evolution that controls the 'instinct' of AI's literary style.
The em-dash controversy is not a UI or convenience issue
To me, the conclusion of the em-dash controversy is understood as an event that fundamentally shakes the ecosystem of writing as a cultural technology.
First. Expanding distrust of AI writing detection.
Turnitin, GPTZero and others already had accuracy controversies, and the em-dash was one of the few remaining weak clues — now even that has disappeared. The AI detection market has come to stand at a point where technological redefinition is inevitable.
Second. Collapse of standards in educational settings.
The era of judging whether AI was used based on style has ended. Teachers and professors say the existing method of finding "traces of AI" in student writing is no longer valid. The method of distinguishing by structure, logic, and style of writing has already shown its limits. The educational world must redesign new evaluation criteria.
Third. Acceleration of the 'undetectable AI writing' debate.
Advertising, content marketing, SNS, political messages — in all these areas, it has become much easier to hide AI usage. This is the collision point between norms demanding transparency and markets pursuing efficiency. The conflict between transparency and efficiency will become more acute.
Regulation also becomes more complex. The AI labeling obligation policies currently being discussed will also become more complex after the em-dash removal. The issue of imitating specific writers' styles may reignite, and questions about what guarantees authenticity will resurface. AI is now erasing its own fingerprints. And humans are left in a world where they must once again think about what writing is.
