AX Field Notes

All casesNo. 064

Chosun Ilbo Replaces Its Copy Desk with an AI Copyeditor

Reporters now run the checker; correction notices rose, while the paper says typos fell

Case no.
064
Primary source
Korean trade press reports (Journalists Association of Korea)
Source published
2026-01-22

1. Why this case

Chosun Ilbo is a Seoul-based national daily. In February 2025 it built and introduced a copyediting AI, developed jointly with the AI company Upstage. By the end of 2025 it had no copy editors. The outsourcing contract with Geuljigi, the subsidiary that did its copyediting, ended in November 2025. By the trade press account published in January 2026, the copyediting team was abolished and AI copyediting adopted across the paper in December.

Many AI cases are told by the company or its vendor, with one set of results. This one was reported a month after the switch by Gija Hyeophoebo (기자협회보), the newspaper of the Journalists Association of Korea, and it carries three readings that do not agree. The union surveyed its members. The trade paper counted the paper’s published corrections. The company said its own review found fewer typos than in the copy-desk period. Each measures something different, and none of them settles the question alone.

The case is also clear about where the work went. Copyediting did not disappear. The second reader between the reporter and the printed page was removed, and the check became a tool that reporters and desk editors run on their own copy before deadline. That shift, and its cost in time and responsibility, is what the union survey and the reporters’ comments describe.


2. Problem

Before the change, copyediting at Chosun Ilbo was a separate human step. The work was outsourced to Geuljigi (글지기), a subsidiary. Before an article was set on the daejang, the proof sheet printed after typesetting to check against the copy, the copyediting team did the typo and fact checks. Reporters could talk to the copy editors about their stories. One Chosun reporter described the division of labor this way (translated from Korean): “The copy desk fixed spelling right away, but when something didn’t make sense on reading, we talked it through.”

The company frames the copyediting AI as one part of a wider plan. In a statement to the trade paper on 2026-01-19, Chosun called it “one part of our ‘AI and digital strategy,’ which applies AI technology to content production and distribution” (translated). The trade paper reports that when Chosun built the system with Upstage in February 2025, it did so as part of a plan to bring AI technology into the whole news service, from article production to distribution.

The company also says the change saved money. In June 2026, answering the union, it said it was reinvesting “the personnel costs saved by introducing AI copyediting” (translated) in translation services (see §4.4).

Confirmed facts: copyediting outsourced to subsidiary Geuljigi; copy desk did typo and fact checks before articles were set on the proof sheet; reporters talked to copy editors about unclear passages; company calls the AI copyediting part of its “AI and digital strategy” (statement of 2026-01-19); the company says AI copyediting saved personnel costs (June 2026); no headcount or cost figure is published.


3. Solution

3.1 From a tool beside the copy desk to no copy desk

Chosun Ilbo built and introduced its copyediting AI system in February 2025 through joint development with Upstage. From February to November 2025, the tool and the copy desk existed side by side.

From November 2025, Chosun Ilbo had no copy editors. The paper ended its outsourcing contract with Geuljigi when the contract expired. According to the January 2026 report, the copyediting team was abolished and AI copyediting fully adopted in December 2025. A later report by the same paper, on 2026-05-13, dates the full adoption and the end of the Geuljigi contract to November 2025.

3.2 The new workflow: writers check their own copy

The typo and fact checks the copy desk used to do before an article reached the proof sheet are now done by reporters and desk editors with the AI checker. Working through a story with a copy editor is no longer possible.

Reporters say they run the checker more than once. One Chosun reporter said (translated): “It sometimes takes longer, too. I don’t run the checker just once but two or three times, because, given the nature of AI, re-running it sometimes finds things it missed the first time.” There is a manual for the tool. The trade paper says that in the first month of full adoption reporters had not yet mastered it.

3.3 Tuning after rollout

In its January 2026 reply, Chosun said (translated): “We are developing a chatbot for ‘Korean transcription of foreign words’ and other matters, and are actively reflecting copyediting issues that arise in content production into the AI and updating its performance. Given the nature of AI, work to improve performance is ongoing.”

Confirmed facts: copyediting AI built and introduced in February 2025 through joint development with Upstage; no copy editors from November 2025; team abolished and full adoption in December 2025 (January report) or November 2025 (May report); reporters and desk editors run the checker before the proof sheet; reporters run it two or three times; a manual exists; a foreign-word transcription chatbot is in development.


4. Impact

Three sources speak to the result, and each measures something different.

4.1 Union survey (newsletter of 2026-01-15)

Chosun’s union reported in its newsletter that “complaints keep coming that people feel confusion and fatigue in writing and closing stories since AI copyediting was fully introduced” (translated), and published a survey of its members on AI copyediting.

Survey itemShare of respondents
Dissatisfied with AI copyediting42.6%
Workload actually increased70.3%
Frequency of typos in print increased79.6%

These are shares of union respondents. The number of respondents was not reported. The figures measure members’ views, not counted errors.

The newsletter also quoted members (translated): “The AI tells us to change ‘12·3 martial law’ to ‘12·12 martial law’,” and “Typos happen even when you check many times, but it feels like all the responsibility falls back on the reporter.” In Korean media, “12·3 martial law” refers to the martial law declared on December 3, 2024. December 12 is the date of the 1979 military coup.

4.2 Correction notices, counted by the trade paper

Gija Hyeophoebo counted the correction notices (“바로잡습니다”) posted on Chosun Ilbo’s website.

Month20242025
September—11
October—14
November820
December729

The paper added two caveats. The criteria for posting a correction can change with the newsroom’s judgment. And in the first month of full adoption, reporters had not yet mastered the tool’s manual, which may have affected the count. November is also the month the copy editors left, by both reports’ timing.

4.3 The company’s reply

Chosun Ilbo answered the newsletter (translated): “The newsletter said ‘the frequency of typos in print has increased significantly since AI copyediting was introduced,’ but that is not true. Our review found it has, if anything, decreased somewhat compared with the period when the copyediting team handled copyediting.” The company did not publish figures or a method for that review. The reply addresses the union’s claim about typo frequency. It does not address the trade paper’s count of correction notices.

4.4 Cost and where the savings went

No source gives the size of the former copy desk or the cost saved. At a 2026-06-02 meeting, the union asked management (translated): “Isn’t the company’s digital strategy limited to headcount cuts and restructuring, with the long-term vision gone missing?” The company answered that it was “investing the personnel costs saved by introducing AI copyediting in improving English translation and in expanding our translation service into major Southeast Asian languages such as Malay, Thai and Indonesian” (translated). No amounts were given.


5. Insight

The check moved to the writer. The copy desk was a second reader who saw the story after the reporter and before the page. Now the reporter and desk editor run a tool on their own copy before deadline. The time cost is no longer a line in a contract with a subsidiary. It shows up as reporters’ deadline time, and 70.3% of union respondents said their workload went up.

Spelling transferred; the conversation did not. The reporter’s distinction is precise: the copy desk fixed spelling, but for passages that did not make sense, reporter and editor talked. The new flow has no step for that part of the copy editor’s work.

A checker that changes its answer needs a human to judge each suggestion. Reporters run the tool two or three times because each run can surface something new. And the tool can be confidently wrong, as with the suggestion to change “12·3” to “12·12.” Every suggestion still needs a reader who knows the story. That reader is now the same person who wrote it.

Accountability went with the work. With no copy desk, an error in print traces back to the reporter alone. The union comment “all the responsibility falls back on the reporter” is a change in who answers for a typo, not only in who catches it.

There is no shared measure. The union measured perception, the trade paper counted notices, and the company cited an unpublished review. Replacing a quality function without a before-and-after measure that both sides accept, such as print errors per page counted the same way in the copy-desk period and after, leaves each side with its own number.

Context. The same report describes Hankyung (Korea Economic Daily), which uses its own copyediting AI. In early January 2026 it moved to shift copy editors to editing roles, then withdrew the plan after staff raised concerns. At a union committee meeting on January 14, the concerns reported included whether AI could check graphics against article text and take over fact-checking at a paper full of numbers.

Limits. All the outcome evidence covers the first weeks after the copy desk left, and all of it comes from reporting by a journalists’ organization. The survey’s sample size is not given. The correction count depends on posting criteria that can change. The company’s counter-finding has no published numbers.

One-line takeaway: Replacing a copy desk with an AI checker moves both the checking and the blame to the reporter. Whether quality held is still disputed because no one counted errors the same way before and after.


6. Implementation hypotheses (not in the source)

6.1 A check the writer runs, not a gate the story must pass

Basis: Reporters and desk editors run the AI checker themselves before the proof sheet; reporters run it two or three times; there is a manual, which reporters had not mastered in the first month.

Because the reporter starts the check rather than the copy desk receiving every story, the checker most likely works on request: a reporter submits a story’s text, gets marked-up suggestions back, and accepts or rejects them one by one. In that setup the production system probably has no record of whether a story was checked, how many times, or which suggestions were accepted. Nothing stops an unchecked story from reaching the proof sheet, so the check is a habit set by the manual, not a step the system enforces. This would also explain why the company and the union could hold different views of quality: the data that would settle it, checks run and suggestions accepted per story, would not be logged where the paper’s production records live.

6.2 Learned from text written before the newest terms

Basis: The system was built and introduced in February 2025; the AI suggested changing “12·3 martial law” to “12·12 martial law”; re-running it finds errors missed the first time; a reporter says it “can’t read that kind of context.”

A copyediting model learns what corrected text looks like from earlier writing. “12·3 martial law” only came into use after December 3, 2024, about two months before the system was introduced, so it was probably rare or absent in whatever text the model learned from, while “12·12” is a long-established date in Korean news writing. A model that has learned the familiar pattern would push new usage toward it. Corrected text also records what changed, not the reporter’s intent or the conversation that resolved an unclear passage, which fits the reporter’s comment about context. Different results on a second run fit a model that generates corrected text rather than applying a fixed rule list, and possibly one that processes a long article in pieces, so each pass sees slightly different context.

6.3 A feedback loop from the newsroom into rules and retraining

Basis: The company says it is “actively reflecting copyediting issues that arise in content production into the AI and updating its performance” and is building a separate chatbot for Korean transcription of foreign words.

The likely loop is a channel where reporters report bad suggestions, a small team that sorts them, and two kinds of fix. Errors about specific terms, such as new events, names, and house spellings, go into a term list that is checked before or after the model runs, since that can be updated the same day. Recurring pattern errors become new training pairs for the next model update. Foreign-word transcription would be split into its own chatbot because it is a lookup problem, matching a foreign name against an agreed spelling, not a grammar problem. Answering it from a maintained list is more reliable than asking a generative checker to recall it.


7. Sources

  1. Gija Hyeophoebo (기자협회보, Journalists Association of Korea), Park Ji-eun — 사람 대신 AI에 신문 교열 맡겼는데, 오탈자 더 늘었다? (We handed newspaper copyediting to AI instead of people, and typos rose?)
    https://journalist.or.kr/news/article.html?no=60175
    2026-01-22 (accessed 2026-11-17). Korean-language article; quotes translated from Korean.
  2. Gija Hyeophoebo (기자협회보), Park Ji-eun — 조선, ‘조근 당번제’ 시도 사과했지만… 디지털 전략은 안갯속? (Chosun apologizes for early-shift rota attempt, but its digital strategy is still unclear)
    https://journalist.or.kr/news/article.html?no=61034
    2026-06-08 (accessed 2026-11-17). Korean-language article; quotes translated from Korean.
  3. Gija Hyeophoebo (기자협회보), Park Ji-eun — 조선NS, 5년 만에 청산 수순… 기자들 ‘위장도급’ 법적대응하나 (Chosun NS heads for liquidation after five years)
    https://journalist.or.kr/news/article.html?no=60877
    2026-05-13 (accessed 2026-11-17). Korean-language article.