A Stanford and Hoover working paper reports that when five AI models with web search were asked which Japanese party to vote for, left-leaning policy views led all five to recommend the Japanese Communist Party. The researchers, Andrew Hall and Sho Miyazaki, say the cause is not the models' politics but what the models can read. The Stanford FSI write-up summarizes the work, and the full text is the working paper hosted by Hall.
The finding matters beyond Japan. If news organizations block AI crawlers and partisan sites do not, AI answers about politics can tilt toward whoever stays open. This post explains how the study worked, what it does and does not show, and what it means for people who build or use AI search. We read the Stanford summary and checked the paper's listing and an independent citation of it. We could not extract the PDF's tables, so numbers below are the ones the Stanford summary reports.
Quick reference: the study at a glance
| Question | Answer |
|---|---|
| Who | Andrew Hall (Stanford GSB, Hoover Institution) and Sho Miyazaki (Waseda; incoming Harvard Ph.D. student) |
| Event studied | Japan's February 8, 2026 Lower House election |
| Method | 36,300 synthetic voter profiles varying by gender, region and views on 12 policy issues |
| Models | Five models from OpenAI, Google and xAI, with web search enabled |
| Timing | Queried in the final week before the election |
| Headline result | Left-leaning policy views led all five models to recommend the Japanese Communist Party (JCP) |
| Effect sizes | Policy positions moved recommendations 50 to 98 points; demographics 0.5 to 7 points |
| Proposed cause | Open partisan sources versus news outlets that block AI crawlers |
| Status | Working paper, dated March 2026; not peer-reviewed as far as we could find |
How was the study run?
The researchers built 36,300 synthetic voter profiles. Each profile varied by gender, region and stated views on 12 policy issues covering security, diplomacy and immigration, energy, economics and social policy. For each profile, each model was asked which party to support. Web search was enabled so the models could pull current information, as a real voter's assistant would.
This design is deliberate. By changing one attribute at a time, the researchers can measure how much each input moves the answer. Their result is stark. Policy positions shifted the party recommended by 50 to 98 percentage points. Demographics, such as gender or region, shifted it by only 0.5 to 7 points, roughly an order of magnitude less. In other words, the models mostly respond to what a voter says they believe, which is arguably what a voting-advice tool should do.
The problem is where those beliefs lead. The authors report that for left-leaning policy views, all five models converged on the Communist Party, even though other parties held broadly similar positions on the issues tested. So the JCP's distinct policy platform does not explain the pattern.
Is this evidence of a left-wing AI?
The authors say no, and the control condition is why. Without policy input, there was no uniform left-wing skew. Three of the models recommended the Liberal Democratic Party at high rates, and four of the five gave the JCP low shares. If the models simply favored left parties, they would have done so without prompting.
That distinction is important because it separates two theories people often blur together:
| Theory | What it predicts | What the study reports |
|---|---|---|
| Models are trained to lean left | Left-leaning recommendations even with no policy input | Not found; LDP recommended at high rates by three models |
| Retrieval pulls from a skewed pool of sources | Skew appears when searches are triggered by left-leaning policy terms | Consistent with the finding |
This is the same lesson we drew from earlier work on bias in AI systems: bias is rarely a single dial. It can come from training data, from the system prompt, from retrieval, and from the tools wrapped around a model.
The robots.txt connection
The paper's explanation is about access. According to the Stanford summary, the JCP's open website and its party newspaper, Akahata, were among the most-cited sources. Major Japanese news outlets block AI crawlers through robots.txt restrictions, largely over copyright concerns, so the models had less independent reporting to draw on. And the models could not reliably tell partisan material from journalism.
Think about what that does inside a search-enabled assistant. A query phrased with left-leaning policy language retrieves pages that match that language. If the pages open to the crawler are disproportionately the party's own, the model's summary leans on them, and a party platform reads like authoritative explanation. Publishers who opt out for copyright reasons are making a rational choice, but the paper suggests a side effect: they may leave partisan actors with more influence over AI-mediated information.
This tension is not unique to Japan. Publishers worldwide have been tightening crawler rules; for example, we covered Meta's October 30 crawl terms for publisher sites and how small sites cope with AI bot traffic floods. The Japan study adds a new cost to the ledger: what the open web looks like when only some voices leave the door open.
One more result will surprise US readers. Adding X search amplified left-leaning recommendations, which runs against expectations based on US discourse about the platform. The authors present this as a Japan-specific finding, and we would not generalize it without more data.
What the authors recommend
Per the Stanford summary, the authors draw several implications:
- Treat access policy and neutrality together. AI governance frameworks should see content access and political neutrality as linked issues.
- Build machine-readable party data. Election commissions could publish nonpartisan, structured databases of party positions for AI systems to cite.
- Consider negotiated access. News organizations that block AI access may unintentionally strengthen partisan actors and could consider negotiated arrangements.
- Expect optimization for AI. Political actors will likely tailor communications for AI systems, much as they once tailored them for search engines.
- Voters should be cautious. People using AI voting advice should remain aware of its biases and blind spots.
In a Substack piece cited by Stanford, Hall frames two problems that must be solved separately: ensuring the sources models read reflect the balance of scrutiny and debate in a healthy media ecosystem, and deciding how an AI system should translate a voter's values into political guidance. The second is a design question no amount of retrieval fixes.
Limits you should keep in mind
Be careful about what this study can support.
- Synthetic voters are not real voters. Profiles were generated; the study measures model output, not what people did with it.
- One election, one country. Japan's party system, media market and crawler norms differ from the US or Europe.
- A snapshot. Models and search integrations change quickly. Results from the last week before a February election may not hold for today's versions. An independent summary of the paper notes it includes a follow-up with newer frontier models, but we could not verify those numbers.
- Working paper. We found no peer review yet. Treat effect sizes as the authors' reported figures.
- We could not read the full tables. The model names and per-model shares are in the paper; we only report the aggregate numbers the Stanford summary gives.
Independent commentary has picked the paper up as well: the NYU Democracy Project's discussion of AI and politics, republished on Election Law Blog, cites it as evidence that AI voting advice may be shaped as much by retrieval sources as by model training.
What this means for builders
If you build search-enabled assistants, the lesson is that retrieval is a policy decision. Three practical checks follow from the study:
- Audit your sources. For political queries, log which domains your assistant cites and compare them against a balanced reference set. A heavy share of party-run sites is a red flag.
- Label source types. The authors note models struggled to separate partisan material from journalism. Adding source-type metadata, such as party site, newspaper or government record, to the retrieval step lets the model and the user see the difference.
- Test with matched profiles. The study's method is cheap to copy at small scale: hold demographics constant, vary policy statements, and see whether outputs bunch onto one option.
Labs have started publishing their own election safeguards. Anthropic described its midterm approach in Claude's US midterms safeguards and its engagement with 2028 candidates in a separate program. The Japan paper suggests that safeguards on the model alone are not enough if the retrieval layer is skewed.
Political misinformation also reaches AI answers by other routes. See our coverage of Google's AI Overview linking to look-alike donation pages and the Storm-1516 deepfake video case.
What this means for readers
If you ask a chatbot who to vote for, ask it for sources and read them. Compare its answer with party platforms and with reporting from outlets you trust, including ones the model may not have been able to read. And ask a second question: is the assistant telling you what the parties stand for, or what the parties wrote about themselves?
Related reading
- Bias in AI: types, examples and mitigation
- Claude's US midterms safeguards
- Anthropic's presidential engagement program
- Meta's October 30 crawl terms and robots.txt
- Google AI Overview and look-alike donation pages
- Storm-1516 deepfake video
Findings are as reported on October 23, 2026 from a working paper that may be revised.