Partly True
SEOUL, Sept. 18 — A claim spreading across Korean social media and online communities this week alleges that OpenAI "stole" 10 million articles from news organizations to develop ChatGPT, framing the chatbot's training data as content taken from publishers without consent. To test the assertion, this newspaper reviewed four primary sources. The review found that the claim pairs a substantiated core — OpenAI's use of news publisher content in model development — with a specific figure and a criminal framing that the source material does not establish.
The claim, worded in Korean as OpenAI having "훔쳤다" — stole — the articles, states that the company obtained 10 million news articles and used them as training material for ChatGPT. As phrased, it asserts two things beyond the existence of OpenAI's training practices: an exact volume of 10 million articles, and that the taking was theft in a settled, factual sense rather than a contested question of data use.
The four primary sources examined support the premise of the claim: OpenAI's development of its models involved the use of news publisher content, and that practice has been a point of dispute. What the sources do not contain is any confirmation of the 10-million-article count. None of the material reviewed traces the figure to a document, dataset disclosure, or filing that would allow the number to be verified.
The second gap is the verb "stole." Describing the conduct as theft imports a legal conclusion, converting what is in substance a contested dispute over training data and publishers' rights into an established wrongdoing. The reviewed sources support that a dispute exists; they do not support presenting the accusation as a proven fact. The distinction matters for readers weighing the claim, because the factual foundation and the accusation built on it are of different evidentiary strength.
On the basis of the four primary sources reviewed, the claim that OpenAI stole 10 million news articles to develop ChatGPT is rated Partly True: the underlying use of news content is real and disputed, while the exact count and the charge of theft remain unproven.
