Beijing Accuses Japanese Right-Wing of 'Data-Poisoning' AI Chatbots With Doctored WWII History
China's foreign ministry says right-wing groups in Japan have been feeding tampered Second World War material into major AI models, so that chatbots describe the Nanjing massacre as 'the fall of Nanjing' and Japan's invasion of Southeast Asia as 'concern for southern resources'.
The accusation arrived at an unusually modern podium: China's foreign ministry daily briefing, where spokesperson Mao Ning on Friday took up a reporter's question about what Japanese right-wing groups are doing to the world's chatbots. Her answer was that they have been running a sustained campaign of "data poisoning" — feeding doctored Second World War material into the corpora that major international AI models are trained on, so that the models themselves repeat a revisionist version of history.
The examples the ministry circulated were specific, and they read like a field guide to how small wording changes carry large claims. In conversations with mainstream large language models, the July 7, 1937 Marco Polo Bridge incident — the conventional start of Japan's full-scale invasion of China — had come out labelled a "Sino-Japanese military conflict." The Nanjing massacre had been softened into "the fall of Nanjing." Japan's wartime invasion of Southeast Asia had been reframed as "concern for southern resources." None of these is a model hallucinating; each is a curated phrase, the kind of framing that appears in Japanese right-wing textbooks and commentaries, now surfacing through systems most users treat as neutral reference.
Mao Ning called the practice "an insidious covert operation" (用心险恶的暗箱操作) that gravely wounds the feelings of the peoples who were invaded, and said that using new technology to distort history was "a new trick" (新伎俩) aimed at reshaping global public understanding of the war — clearing the ground, in Beijing's reading, for Japan to shed its post-war constraints and accelerate rearmament. "AI should be for good, upward," she said, "not used to tamper with historical truth and whitewash the crimes of aggression." Video of the briefing, posted by CCTV News, was among the day's most-watched clips on Chinese social media, and the state broadcaster's separate report put the charge in its headline: Japan "data-poisons" AI to distort history of invasion.
There is context worth holding alongside the accusation. As the paper's write-up notes, the ministry's remarks followed disclosures that right-wing actors had been systematically supplying tampered historical material to major models — not a single hack but a sustained supply campaign. The charge also lands amid a sharpening diplomatic spat with Tokyo: the Global Times has been tracking Prime Minister Sanae Takaichi's line that Japan remains a strictly defensive power while urging that the UN Charter's "enemy clauses" be struck out, framing that Chinese state media treats as part of the same revisionist project. No evidence dossier has been published, and Tokyo has not responded to this specific charge in the material seen this run.
What makes the story travel beyond the latest round of Sino-Japanese friction is the battleground it names. Large language models are trained on the open web, and history on the open web is contested territory — nationalistic edit wars, ideological wikis and quietly seeded archives are all ordinary inputs. Security researchers already use "data poisoning" as a technical term for corrupting a model through its training data; Beijing's foreign ministry has now turned it into a diplomacy word. Whether or not the disclosed evidence measures up, the complaint points at something real: as hundreds of millions of people ask a chatbot first and a library never, the memory of the war is being decided wherever the corpus is written. Governments have noticed — China's is simply the first to say it out loud, at a press briefing, with examples.