Japan’s “Japanese Language Shrinkage” Warning, Why Domestic AI Options Now Matter

 Daisuke Okanohara (Preferred Networks co-founder; appointed President/Representative Director on November 26, 2025) is an influential voice in Japan’s AI debate, including as a member of the “AI Business Operator Guidelines” expert committee. 株式会社Preferred Networks+1 He argues that global competition (Google, OpenAI, Chinese players, etc.) will keep improving models on a 6–12 month cadence, so we should focus on the underlying trend rather than each announcement. He frames three risks of not having a domestic (Japanese) AI option: (1) an economic risk—AI will become infrastructure and a platform for new industries, so lacking domestic choices matters over the long term; (2) a language-and-culture risk—overseas frontier models may speak Japanese well, but the share of Japanese in their training data can be tiny (e.g., 0.5% or even 0.1%), which can cause generated Japanese to converge on a narrow set of expressions and “thin out” the language; and (3) a sovereignty/security risk—within 5–10 years AI will move into critical domains (government and corporate data, health data), and relying only on foreign AI could reduce controllability if relations worsen. He says Japan does not need to make everything domestic, but must secure domestic choices for the areas it wants to control, combining data partnerships (e.g., public Japanese-language resources) with efficiency under Japan’s power constraints, while acknowledging that high-quality, safely usable data remains a major bottleneck.

By naming “linguistic convergence” as a concrete risk, he reframes domestic AI as cultural infrastructure—not just a race for benchmark scores.
But because data building and ROI are long-horizon, success depends on a two-layer strategy: public-interest domains (language/critical data) plus business differentiation.