AI voting advice in Japan reportedly concentrated on the Communist Party for left-leaning profiles

Comparing recommendations across 36,300 synthetic profiles

Researchers created 36,300 synthetic voter profiles and asked five AI models, with web search enabled, which Japanese party each profile should support. They report that policy positions had much larger effects on recommendations than demographic factors, and that left-leaning profiles were overwhelmingly directed to the JCP.

In the study’s control condition without policy positions, four of the five models had low JCP recommendation shares, while three recommended the Liberal Democratic Party at high rates.

The researchers say that adding left-leaning policy positions made all five tested models converge strongly on recommending the JCP, a pattern they attribute to the political information available through web search.

The reported result suggests that AI political recommendations can be shaped by which sources are accessible to web search, not only by a model’s training. The researchers say the JCP’s openly accessible party newspaper and website were frequently cited, while major Japanese news outlets restricted AI crawler access.