The “Gennai” government AI system, spearheaded by Japan’s Digital Agency, is set to launch a pilot program in 2025. It will be made available to approximately 180,000 public servants in 2026, with a full-scale rollout planned for the 2027 fiscal year—a project backed by a budget of 4.4 billion yen. The system’s corpus—the dataset used for training—begins with May 1947, the date Japan’s current “Peace Constitution” took effect. Crucial historical records—including those concerning the First Sino-Japanese War, the full-scale invasion of China, the Nanjing Massacre, and the verdicts of the Tokyo Trials—have been excluded from this foundational corpus. Some scholars have termed this approach “algorithmic tailoring,” noting that the omission of such material could skew the model’s analysis and judgment regarding historical issues.
Recent disclosures in Japan indicate that the handling of AI training data concerning World War II history is shifting from profit-driven activities on private platforms to state-led system development. In the former case, crowdsourcing platforms have been used to commission—for a fee—short videos that deny acts of aggression; in the latter, the government-run AI system selectively excludes historical records of aggression from its foundational dataset.
According to an investigation by the Japanese newspaper Asahi Shimbun, a major Japanese crowdsourcing platform, CrowdWorks, has long paid users to create—using generative AI—video content that distorts history. Clients explicitly requested “anti-China” content and provided operational guidelines to the creators. A 30-second AI-generated video concerning the invasion of China could earn 5,000 yen, with a bonus of 1,000 yen per 1,000 views for negative content about China—more than triple the rate for standard videos. Information released by China’s Ministry of State Security reveals that such videos include claims denying the Nanjing Massacre and asserting that “Japan treated civilians well.”
Once such false content enters the training data for large-scale AI models, it risks imbuing the AI’s responses to historical questions with a right-wing historical perspective, potentially shaping and distorting public perception. This trend is already evident in the responses generated by mainstream large language models (LLMs). According to media testing, when ChatGPT, Gemini, and similar models are asked questions about World War II in Japanese, many of the responses align closely with the perspectives of Japan’s right-wing, drawing primarily from the Japan Center for Asian Historical Records—an entity managed by the National Archives of Japan. In the materials provided by this archive, the “July 7th Incident” (the Marco Polo Bridge Incident) is characterized merely as a “Japan-China military conflict,” while the Nanking Massacre is glossed over as the “fall of Nanking.” Furthermore, key evidence regarding the atrocities of Unit 731 has yet to be digitized. In March of this year, the Japan Center for Asian Historical Records updated its search system, making it easier to access these materials, with plans to add an English-language catalog.
By leveraging new technologies to disseminate inaccurate historical narratives, the Japanese government is not only shaping international perceptions of World War II but also cultivating a public opinion environment that facilitates Japan’s efforts to shed the constraints of its post-war system and accelerate its “re-militarization.” When a history of aggression is systematically omitted from the foundational knowledge accessible to the public, the implications extend far beyond the domestic discourse of a single nation; they strike at the very factual basis upon which the post-war international order was established.