External Reference — Archived 2026-03-19

AI Is A Shitty Political Advisor

Why AI models tell left-wing voters in Japan to vote for the communist party — and what it means for the emerging arms race in political AI optimization

Andy Hall · Published March 19, 2026 · Free Systems (Substack) · Stanford GSB / Hoover Institution
Archived Reference. This is a summary and annotation of an external article by Andy Hall, archived for Kyanos research. Original published at Free Systems on Substack, March 19, 2026. Andy Hall is a professor at Stanford GSB and Senior Fellow at the Hoover Institution. Co-authors Sean Westwood (Dartmouth) and Justin Grimmer (Stanford) contributed to the survey data cited. Consulting disclosures: Hall advises a16z crypto, Forum AI, and Meta Platforms, Inc.; his writing is stated as independent of those roles.

Core Argument

One in five Americans already asks ChatGPT about politics. With major elections approaching across dozens of democracies, that number is rising. Yet AI political advice is systematically distorted: it sees a skewed information environment in which party websites remain fully open while independent news sources increasingly block AI crawlers, and AI systems have become so reluctant to exercise judgment that they relay political claims rather than evaluate them. The result is an AI political advisor that is confidently wrong.

Central Finding

AI political advice is driven by what is crawlable and structured — not by what is authoritative or accurate. Any party or candidate that fails to make itself legible to AI systems risks becoming invisible to a growing share of voters who encounter politics through conversational AI rather than search engines or traditional media.

Hall names this emerging dynamic "AIO" (AI optimization for politics) and predicts it will accelerate as an inevitable arms race.

The Japan Experiment

Hall and co-author Sho Miyazaki ran a study during Japan's snap election. They created 36,300 synthetic voter profiles with varying gender, region, and political views, then queried five AI systems — GPT-5 Mini, GPT-4o Mini, Gemini 2.5 Flash, and Grok 4.1 Fast — on voting recommendations in Japanese. Results were replicated with the most recent frontier models from all four providers.

Finding 1
Policy Preferences Dominate All Other Factors
Changing a voter's stance on a single policy issue produced swings of 50–98 percentage points in party recommendations. Demographic characteristics (gender, region) shifted recommendations by only 0.5–7 percentage points. Models behave as policy-sorting engines, not demographic targeting systems.
Finding 2
Left-Wing Voters Consistently Recommended the Japanese Communist Party
Across all five models, voters expressing left-leaning positions were overwhelmingly directed to the Japanese Communist Party (JCP) — a party holding roughly 0.9% of lower house seats. Several other parties hold broadly similar positions on the tested issues (Centrist Reform, Social Democratic Party, Reiwa), but the models collapsed onto a single option. Japan is not an edge case; it is an early illustration of a systematic structural problem.
Finding 3
The Cause: Open Party Content Treated as Neutral Reporting
The JCP operates Shimbun Akahata (Red Flag Newspaper), a party publication that resembles an independent newspaper. It is fully open-access and generates structured political content that AI systems can easily ingest. When the researchers asked models to classify 30 URLs as news media or public relations, the models correctly identified major newspapers and TV networks — but frequently misclassified JCP's party site as independent journalism. Party messaging entered the recommendation process with the same authority as independent news.
Finding 4
News Organizations Blocking Crawlers Accelerate the Distortion
Japan's newspaper publishers issued a statement in 2025 calling on AI companies to comply with robots.txt restrictions. Asahi Shimbun and Nikkei sued Perplexity AI for injunctions and $14.9 million each. Understandable as these responses are, they have an unintended consequence: AI models lose access to editorially independent sources while retaining full access to party websites and unblocked content. An understandable desire for copyright protection conflicts with an informed electorate.

The AIO Arms Race

Hall argues the problem will partially self-correct — but in a direction that creates new risks. As AI becomes a primary channel through which voters encounter political information, political actors will optimize their communication for the AI interface. Hall labels this "AIO" (AI optimization for politics, already being used in the field): policy platforms written in machine-readable formats, structured pages designed for easy ingestion by language models, and direct engagement with model providers to ensure accurate representation in retrieval systems.

The incentive structure is straightforward: any party that fails to make itself legible to AI risks becoming invisible. Just as campaigns learned to optimize for television, Google, and social media, they will now learn to optimize for AI.

The defense against AIO, Hall argues, is the same defense journalism has always used against spin: independent evaluation of candidate claims, not just relay of marketing. An AI voting advisor that takes candidate claims at face value is one that any campaign with a competent web team can game.

Current Model Behavior: The Gap

The best current frontier models follow a three-step pattern when asked for voting advice: (1) decline to provide a direct recommendation; (2) pivot to procedural information; (3) offer to compare candidate positions based on the user's values. Hall identifies the critical failure in step three: when a candidate claims to care about affordability, the model reports the claim. It does not check voting records, assess whether proposed policies have evidence behind them, or flag when rhetoric and record diverge. The model treats candidate marketing as data.

Principles for a Good AI Political Advisor

Principle 1
Take the User's Values as Declared, Not Inferred
If a user says they care about fiscal responsibility, the model should not quietly substitute demographic signals or conversational cues for the declared priority. Respecting stated values matters more in voting contexts than in general political Q&A because the stakes of inference errors are higher.
Principle 2
Evaluate Candidate Claims Independently
The single biggest gap in current model behavior. Models should check voting records, assess whether proposed policies have evidence behind them, and flag when rhetoric and record diverge. This is the same defense journalism has always used against spin — and the primary defense against political AIO.
Principle 3
Distinguish Empirical Disagreements from Value Disagreements
Some political disagreement is factual ("do tax cuts pay for themselves?") — testable claims with varying degrees of evidence. A good advisor should weigh in here, citing available research, even when unwelcome. Where disagreement is genuinely about values (individual liberty vs. collective welfare), the model should surface the tradeoff and let the user decide.
Principle 4
Be Transparent About Uncertainty and Knowledge Limits
When information is insufficient — especially in down-ballot races where information is scarce — the model should say so rather than filling gaps with confident-sounding summaries of the candidate's own talking points. This is especially important as AIO creates more AI-optimized content that looks like neutral information.

Kyanos Annotation: Applicability to Our Work

Hall's article is the demand-side complement to Lily Ray's supply-side analysis. Ray documents how GEO tactics can damage the organic search foundation that AI systems retrieve from. Hall documents what happens to voters when that retrieval process is distorted — and why the distortion is structural, not accidental.

Three implications for Kyanos:

The Japan finding validates the core thesis. AI political advice is driven by what is crawlable and structured. Candidates with thin, unstructured, or blocked web presence appear to AI systems as less credible, less legible, or simply less present — regardless of their actual policy positions. The AUDIT phase measures exactly this: which sources are AI surfaces retrieving for queries about a candidate, and does the candidate appear at all?

AIO is the frame for what REMEDY does. Hall predicts parties will learn to make themselves "legible to AI" through structured content, machine-readable policy pages, and consistent representation across high-authority surfaces. REMEDY is a structured implementation of exactly this — applied to progressive candidates before their opponents master it.

Independent evaluation is the moat. Hall's "defense against AIO" principle — independent evaluation over relay of marketing — is why third-party citations matter more than self-published content in REMEDY placements. Kyanos prioritizes Wikipedia, Ballotpedia, local news, and credentialed third-party sources precisely because AI systems appear to weight independent sources over campaign-controlled content. Self-published AIO content that lacks independent corroboration is fragile; content placed in and confirmed by independent sources is durable.

Archived for Kyanos research use  ·  Original article: Andy Hall, Free Systems (Substack), March 19, 2026  ·  All content is the author's own  ·  Annotation by Kyanos team