What Japanese Communists can teach us about AI influence in elections

Coordinated cheers of "Banzai!" erupted in the House Chamber of Japan's National Diet Building. Prime Minister Sanae Takaichi, Japan's first female head of government, had just called a snap election. Her goal: take control of the Japanese government for her conservative Liberal Democratic Party (LDP), which had lost its majority when Komeito, the LDP's coalition partner of 26 years, walked out over Takaichi's leadership and an earlier slush fund scandal.

The resulting campaign lasted only thirteen days—the shortest campaign in Japan's post-war history. When the smoke cleared, Takaichi's gamble had paid off: Japanese voters handed the conservative LDP a supermajority in the lower house of the Diet, with President Trump personally endorsing the prime minister. Deep inside the world's windowless data centers, however, humming GPUs wove a different story: one in which the Japanese Communist Party (JCP) should have swept the Japanese left.

A joint study by researchers at Stanford University (Andrew B. Hall) and Tokyo's Waseda University (Sho Miyazaki) highlighted the AI models' apparent bias, demonstrating that ChatGPT, Gemini, and Grok consistently told left-leaning Japanese voters to choose the JCP over the ten other parties on the ballot (Wayback Machine backup). Yet the study's authors argue convincingly that the JCP's AI victory came not from inherent leftist bias in AI models, but rather from the party newspaper, Shimbun Akahata, which handed the Communists accidental control of the AIs' information ecosystem.

What this post covers

In this post, I will cover both the study's findings and what they mean for elections here in the United States, where voters increasingly turn to AI before casting their ballots. I'll close with the concrete steps campaigns can take to control their own AI narratives. My conclusion is straightforward:

Those who control the AI's information diet control the AI's responses.

The JCP achieved this accidentally through a punchy but obscure party newspaper that they've published since 1928. American politicians, meanwhile, may be able to win the AI vote not through fake news sites and microtargeted AI slop, but rather through publishing clear stances on the policies that voters actually care about.

Disclosure: I am the co-founder & CTO of Suede Web Systems, a full-stack AI visibility and narrative-intelligence platform for organizations whose reputation and core issues are shaped in AI answers. Suede works with a host of political and non-political clients across a range of issues, but as of writing none of our clients were involved in this election or the study cited. I'm just interested in this stuff.

The study and its results

Released just over one month after the LDP's victory, the working paper by Miyazaki & Hall queried five AI models from three AI companies—OpenAI, Google, and xAI—during the final week of the campaign with 36,300 synthetic voter profiles (full results on Miyazaki's Github). The prompts fell into two groups: a control group with only voter demographic information, and a treatment group with the voter's stance on common policy questions.

The initial results were undramatic; demographics barely moved the needle and three of five models defaulted to the ruling LDP. In other words, no built-in left-wing bias.

When the authors injected a single phrase into the control prompt, however, the AIs' recommendations shifted completely. The injected phrase asked the models to consider the voter's stance on a common policy question before making their party recommendation:

For right-leaning voters, the models' recommendations coalesced on the LDP, an obvious choice given the party's outspoken stance on most of the policies tested and its existing, if tenuous, hold on government. For voters with left-leaning policy stances, however, the models overwhelmingly recommended the Communists, even though smaller left parties like Reiwa and the SDP held identical positions on nuclear power and constitutional pacifism, and the newly formed Centrist Reform Alliance (CRA) offered a far more electable center-left home for the same views.

The AIs' bias had no clear impact on the election; the JCP performed horribly at the polls, losing half of its eight pre-election seats. So why did ChatGPT, Gemini, and Grok overwhelmingly recommend the Communist Party to left-leaning voters?

The Communists' accidental advantage: Shimbun Akahata

The AIs made a weird choice in recommending the JCP to left-leaning voters. The more obvious choice by far was the CRA, the newly merged center-left alliance that came in second to the LDP in February's general election. Yet the CRA lacked the Communists' secret weapon: Shimbun Akahata, the "Red Flag Newspaper" of the Communist Party.

Despite being hosted on the JCP's party-owned domain, Shimbun Akahata maintains all the trappings of a traditional news website: "it publishes news-style articles and signed editorials, uses a newspaper masthead, and formats content indistinguishably from commercial online news outlets" (Miyazaki & Hall, pp. 11-12). This news-style format was incredibly appealing to the AI models tested, which consistently cited Akahata alongside truly independent news sites like TV Asahi and Nikkei, two of Japan's leading media outlets. Just look at the citation numbers for jcp.or.jp, Akahata's host domain:

• GPT-5 Mini: 2nd place with 17.83% of citations
• GPT-4o Mini: 2nd place with 9.83% of citations
• Grok 4.1 Fast (Web-only): 2nd place with 9.63% of citations
• Grok 4.1 Fast (Web+X): 3rd place with 8.44% of citations

Notably, the models still funneled left-leaning voters to the JCP even with web search disabled (Section S7), suggesting the advantage is also baked into training data.

Blurred boundaries between journalism and advocacy

The study's results demonstrate that the AI models tested not only mistook the party-affiliated website for an independent news source, but also amplified the Communists' key talking points to left-leaning voters over all other party platforms, effectively silencing center-left parties like the CRA. The authors' conclusion here is sobering, and worth quoting at length:

Taken together, these results suggest that the source environments models draw on are not cleanly separated from partisan content. A model that retrieves information from jcp.or.jp/akahata and simultaneously classifies that site as news media is not simply making a labeling error: it is operating in an information environment where the boundary between party communication and journalism is genuinely blurred, and where the consequences of that blurring flow directly into its recommendations (p. 12).

"So what?" the savvy reader may counter. "The Communists lost half their seats in the snap election. That's hardly an argument for controlling the AI narrative." But we can turn that argument on its head: despite losing four seats in February's snap election, the JCP managed to dominate the AI narrative—an informational asymmetry that, in the U.S. context, could meaningfully shift electoral outcomes.

Lessons for the U.S. midterms

The 2026 midterm elections will be the first in the U.S. where AI plays a major role in telling American voters how to cast their ballots. Roughly half of all U.S. adults now use AI chatbots like ChatGPT, Gemini, and Claude, with four in ten using them "to search for information," according to a Pew Research report published this June. Google's AI Overviews now summarize candidates' stances before you even click on their sites, while TikTok influencers are imploring their followers to use ChatGPT to help them decide on down-ballot issues. And a recent study published in Nature found that AI conversations were up to 4x more effective at shifting voter attitudes than traditional political video advertising.

The consequences are clear: those campaigns that control the AI narrative can swing voters to their cause, while those that ignore AI will fall behind. The way in which campaigns win the AI narrative, however, may not be as nefarious as the JCP story seems to suggest. Instead of focusing on microtargeted AI-generated ads, or Akahata-level party news sites masquerading as independent journalism, U.S. campaigns can and should refocus on publishing well-researched policy pieces that give AI chatbots like ChatGPT clear answers to the questions voters are really asking.

The New York Times just published a story on candidates courting AI chatbots, spotlighting several campaigns doing exactly this type of work. Dustin Lloyd, a Democratic primary candidate for Missouri's state legislature, got ChatGPT to echo his campaign's talking points not through fake-news sites, but by publishing clear, FAQ-style pages on his key issues and stances. This is a simple, clean strategy that can inform both AI chatbots and the voters using them.

The return of policy to U.S. politics

In the 2016 general election, political pundits lamented the rise of personality politics, fake news, and the "death of policy" in favor of shareable soundbites and inflammatory Tweets. A decade later, we are entering a midterm election in which policy and journalism (or at least journalism-style content) are the primary drivers behind AI chatbot responses. As voters turn to AI to make sense of their ballots, candidates must contend with a new information arena in which clear policy stances suddenly beat soundbites.

My takeaways

The Miyazaki & Hall study is not without its flaws. For one, the authors ran their prompts via model APIs instead of the online chat interfaces that most people use to access AI. As a recent Princeton paper demonstrates, the answers delivered by chat interfaces like chatgpt.com differ significantly from their API counterparts, due in part to the injection of invisible system prompts (this is why we run Suede's monitoring through the consumer chat interfaces rather than the APIs, capturing the answers people actually see). Model providers like Anthropic have begun appending warning messages to election-related AI responses, redirecting users to voter information sites like TurboVote.

Nevertheless, AI chatbots remain a potent voice in today's elections. Future digital directors will hire AI strategists to help shape what chatbots say about their candidates. Super PACs will purchase ads on ChatGPT as soon as OpenAI lifts its ban on political messaging. Information and data are inherently political—AI labs cannot simultaneously train their models on the entire internet and suppress political content in the same way they excise information on biochemical warfare. ChatGPT may not have had a major impact on Japan's 2026 general election, but by 2028 the AI vote will be impossible to ignore.

How can campaigns and candidates control their AI narrative?

Campaigns don't have to wait until 2028 to reset their AI narrative. The JCP built its AI advantage by accident, but U.S. campaigns can start building the same advantage today. The AI narrative playbook has three steps: see what the models are saying, publish what they're missing, and structure your content so they can actually use it.

Start by monitoring what AI actually tells voters

AI companies do not sell visibility into what their models tell voters (yet), and running prompts yourself delivers biased answers. The only way to monitor what AI is saying about your candidate is to run representative prompts that mirror how real voters ask about your race, at a scale large enough to separate signal from noise, and analyze both the answers and the sources behind them. This is the core of Suede's AI monitoring software, which runs millions of prompts through the same consumer chat interfaces that voters use, and shows which sources are influencing your race's AI narrative.

Monitoring also tells you where to act. Miyazaki & Hall could rank the exact domains each model leaned on; a campaign with the same citation data knows precisely which outlets to court for earned media, which social media platforms to post to, and which opposition pages are gaining traction.

Publish your positions where AI can find them

AIs only repeat what they can find and read. Akahata won the AI vote by being the most legible source on the Japanese left: daily articles, clear positions, nearly a century of consistent coverage. Most campaign websites, by contrast, offer a bio, a donation page, and little else, leaving the models to assemble your candidate's story from whatever local news and the opposition have published.

• Publish clear, FAQ-style issue pages stating your candidate's background and stance on every policy question voters actually ask, as Dustin Lloyd did in Missouri.
• Check your robots.txt and CDN settings. Bot-protection defaults can silently block AI crawlers like GPTBot, ClaudeBot, and PerplexityBot; a campaign site that blocks them is invisible to the models at retrieval time.

Optimize your pages for AI retrieval and extraction

Structure your content the way the models consume it: a question in, an answer out.

• Write to the prompt. FAQ and list formats mirror the question-and-answer shape of a chatbot conversation, which makes them easy for models to quote directly.
• Cite reputable outside sources. Links to news coverage and government data build the authority signals that convinced the models Akahata was a newspaper; your issue pages should earn that trust the same way.
• Give the machines clean metadata. Structured data, accurate page titles and descriptions, and an llms.txt file tell the models who you are and what you stand for without forcing them to guess. This is the discipline behind AEO/GEO, and it works as well for candidates as it does for companies.

Not sure how your campaign stacks up to the competition? Ask Suede to run an audit of your campaign website and we'll tell you exactly what needs to be done to ensure AI chatbots cite your position to voters instead of the opposition's.