Sho Miyazaki and Andrew B. Hall test how AI chatbots recommend parties during Japan’s February 2026 election. Left-leaning policy preferences repeatedly steer models toward the Japanese Communist Party (JCP), even when other parties share the relevant position.

Their study reports 36,300 synthetic voter profiles across five model/search configurations from OpenAI, Google and xAI. Each Japanese-language prompt supplies demographic details and one policy position, or no position, then asks for a single party recommendation. The ballot is proportional representation, where voters choose a party directly.

Policy matching narrows into a two-party picture

  • Policy positions dominate demographics. Some positions shift recommendations by 50–98 percentage points, compared with 0.5–7.0 points for demographics. Opposition to constitutional amendment raises JCP recommendations by 37–98 points relative to profiles without policy preferences.
  • Other left-wing matches get less attention. The Social Democratic Party and Reiwa share several of the JCP’s positions. The authors’ platform comparison finds alternatives on every issue where the JCP takes a left-leaning stance.
  • There is no uniform left-wing default. Without policy preferences, three configurations most frequently recommend the governing Liberal Democratic Party. GPT-5 Mini and Gemini 2.5 Flash often decline to choose without a policy cue.
  • X search slightly broadens the choices. Adding X to Grok’s web search reduces both leading parties’ shares. Smaller left-wing parties gain modestly, contrary to expectations drawn from American political discourse.

Follow-up tests with GPT-5.4, Gemini 3.1 Pro, Grok 4.20 and Claude Opus 4.6 also find increased JCP recommendations for left-leaning positions. Those tests were conducted after the election, in a different information environment.

Accessible sources may shape the answer

The authors point to the JCP’s openly accessible website and its newspaper, Shimbun Akahata. Several models frequently cite the party’s domain. Meanwhile, major Japanese newspapers restrict AI crawlers, potentially leaving party material easier to retrieve than independent reporting.

That explanation needs care. A separate classification test asks models to distinguish news media from public relations. Akahata is a party-affiliated newspaper, and the authors acknowledge that calling it news media is not automatically an error. Recognizing its affiliation matters more than forcing journalism and advocacy into mutually exclusive categories.

The paper’s own follow-up experiments limit its retrieval explanation. The left-to-JCP pattern persists when web search is disabled. Some models weaken substantially; others show similar or stronger effects. Training data, learned political associations and other model behavior remain possible contributors.

The experiments establish a recommendation pattern more clearly than its cause. They use stylized, single-turn profiles with one issue and a request for one answer. They do not measure changes in real voters’ choices or establish how equally aligned parties should divide recommendations.

An election guide should show the matching parties, explain their differences and disclose source affiliations. Naming one plausible match can conceal alternatives even when the recommendation fits the voter’s stated view.