Five AI models told left-leaning Japanese voters to select the Communists

Two researchers tested five AI models from OpenAI, Google and xAI in the last week before the Japan election on 8 February 2026. They gave the models 36,300 made-up voter profiles. When a profile had left-leaning views, all five models named the Japanese Communist Party. The researchers say the cause is the web content that the models can get.

How it works

The researchers, Andrew Hall and Sho Miyazaki, gave each profile a gender, a region and views on 12 policy issues. Each model had web search on and had to name one party.

The numbers

Policy views moved the party selection by 50 to 98 percentage points. Gender and region moved it by 0.5 to 7 points. With no policy input, three of the five models frequently named the Liberal Democratic Party, and four of the five did not name the Communists frequently.

What is not known

The researchers say that large Japanese news sites keep AI crawlers (software that reads web pages) out with robots.txt files, because of copyright rules. The party newspaper Akahata is open, and the models used it frequently. The paper is a working paper. The source does not say that other teams checked it.

Insight: Open sources shape what an AI agent says

The party newspaper was open to the models, and the news sites were closed because of copyright. In an earlier test, a seller added a correct line to an item text. Agents then found incorrect web pages and gave other items a lower rank. This reads as a race over what agents can read: the side that stays open can lead the result. Publishers that keep crawlers out can decrease their influence. Leaders must test what their agents get.