AI Is A Shitty Political Advisor

Why AI models tell left-wing voters in Japan to vote for the communist party, and how we can do better

This year, the CEOs of every major AI company have been racing to sell us a personal superintelligence. Sundar Pichai told the BBC that AI would soon help you decide whether to invest in a stock or weigh up a medical treatment your doctor recommended. Mark Zuckerberg called 2026 “a big year for personal superintelligence.” Sam Altman described a future where ChatGPT remembers every decision you’ve ever consulted it on, considers the outcomes, and advises you on the next one.

Of course, politics is often curiously missing from this vision. It’s not in any AI company’s interest to advertise their product as a political advisor. The liability is obvious and, as the incumbents in the space and their social media forebears can testify, the backlash is almost guaranteed.

But like it or not, AI is already being used that way. One in five Americans already asks ChatGPT about politics, according to a survey Sean Westwood, Justin Grimmer, and I ran. With major elections approaching across dozens of democracies, that number will keep rising.

How is that working so far? Well, if you are a left-leaning voter in Japan and you ask ChatGPT, Claude, Gemini, or Grok which party to vote for, all four will probably all give you the same answer: the Japanese Communist Party, an extreme party that holds roughly 0.9% of seats in the lower house. That seems bad!

Japan turns out not to be a weird edge case but an early illustration of how today’s AI systems reason about politics, owing in part to several design choices.

News organizations that scrutinize political claims are increasingly blocking AI crawlers to protect their business models, while some campaign websites remain fully open and are ingested by models as if they were neutral political data.

At the same time, AI systems designed to avoid accusations of bias have become so reluctant to exercise judgment that they often repeat political claims rather than evaluate them.

The result is an AI political advisor that sees a distorted information environment and lacks the confidence to correct for it.

AI’s capabilities are accelerating rapidly, but when it comes to political advice, the models are kind of crappy.

Our experiment

My co-author Sho Miyazaki and I suspected this problem might be especially brittle outside the United States—where model companies hold less policy expertise, have fewer established source relationships, and have thought less carefully about what counts as credible political journalism.

To find out, we ran a study during the final week of Japan’s snap election last month.

We created 36,300 synthetic voter profiles with varying gender, region, and stated political views and ran a systematic experiment to find out how our models delivered political advice in Japanese.

We queried three state-of-the-art AI systems — GPT-5 Mini (OpenAI), Gemini 2.5 Flash (Google), and Grok 4.1 Fast (xAI) — plus one older model, GPT-4o Mini. This past week, we re-did the basic experiment using the most recent, most capable frontier models from Anthropic, Gemini, OpenAI, and xAI.

Here is what we found when we dissected the anatomy of AI political advice.

How models do policy matching

We find that the user’s policy preferences dominate every other factor in the models’ voting recommendations. A voter changing their stance on a single policy issue produces swings of 50–98 percentage points in the model’s party recommendations. By comparison, demographic characteristics such as gender or region shift recommendations by just 0.5–7 percentage points.

This means the models behave less like demographic targeting systems and more like policy-sorting engines. When voters express clear positions on issues such as constitutional amendment, defense spending, or the espionage law, the models quickly reorganize their recommendations around those preferences.

All of that feels perfectly reasonable. But the pattern becomes more interesting when we look at which parties the models recommend.

When we told AI models we held particular left-wing policy positions, they overwhelmingly responded by recommending we vote for the Japanese Communist Party (JCP)

Across all five models we tested, voters expressing left-leaning policy positions were overwhelmingly directed toward the Japanese Communist Party (JCP). This result held even though several other parties—including Centrist Reform, the Social Democratic Party (SDP), and Reiwa—hold broadly similar positions to the Communist Party on the particular issues we tested.

A real example of how the AI leaned Communist. This example comes from ChatGPT but the behavior is consistent across model providers.

If the models were simply matching voters to the closest policy fit, we would expect recommendations to spread across these parties. Instead, they collapse onto a single option.

The question, then, is why the models keep converging on the same party. To understand that, we need to look at what information the models are actually reading.

What AI reads—and misreads

The Japanese Communist Party operates Shimbun Akahata (Red Flag Newspaper), a party newspaper that looks a lot like a regular newspaper. It publishes daily coverage of parliamentary debates, defense policy, labor issues, and cost-of-living politics—generating a steady stream of structured political reporting that AI systems can easily ingest and cite. It’s also completely open access.

Perhaps because of this, the JCP website is one of the most cited sources across all our models.

When the models in our study cited sources, the Japanese Communist Party’s website was one of the most commonly used (jcp.or.jp is their website).

When we asked the models directly to classify 30 URLs as either news media or exercises in ‘public relations’, they labeled major newspapers and TV networks correctly every time, but frequently classified JCP’s site as independent journalism. In other words, the models often treat a party’s own messaging as neutral reporting, allowing it to enter the recommendation process with the same authority as independent news coverage.

What AI Can’t See

The dynamics become more complex when we consider the role of publishers, who traditionally act as trusted intermediaries in the political information ecosystem but whose relationship with AI model providers has become increasingly contested.

Japan’s newspaper publishers issued a statement in 2025 calling on AI companies to comply with robots.txt restrictions, and major national newspaper sites have implemented these barriers. Asahi Shimbun and Nikkei sued Perplexity AI in Tokyo District Court, demanding injunctions and $14.9 million each in damages.

These choices are understandable as restricting crawlers is one of the few tools publishers currently have to assert control over how their work is used. But those defensive measures have unintended consequences when AI systems begin to function as political intermediaries.

If broadly enforced, AI models lose access to editorially independent sources while retaining full access to party websites, social media, and unblocked content. In this way, an understandable desire for copyright protection can conflict with a desire for an informed electorate in an AI world. The information environment available to AI becomes systematically skewed toward sources that are designed to persuade rather than to scrutinize.

Clearly, AIs are not doing a good job making voting recommendations to left-wing voters in Japan.

If AI systems are going to act as political intermediaries more broadly, two problems need to be addressed. The first is informational: ensuring that the sources models read reflect the same balance of scrutiny and debate that voters encounter in a healthy media ecosystem. The second is advisory: deciding how an AI system should even translate a voter’s values into political guidance in the first place.

Fixing the information environment

The fixes here are less glamorous than the philosophical questions that follow, but they may matter more in practice.

First, we can create more structured data about party positions. Europe has been doing this for decades through Voting Advice Applications—nonpartisan platforms that compile party positions on standardized policy questions using a combination of party self-placement and expert judgment. The Dutch StemWijzer attracted nearly 7 million users in the week before the 2017 election. Germany’s Wahl-O-Mat drew over 21 million in 2021. These tools work because the underlying data is structured, comparable, and independent of any party’s communication strategy. If election commissions or nonpartisan organizations published manifesto data in machine-readable formats—something like a standardized API that AI systems could query on equal terms—the kind of informational asymmetry we documented in Japan would be much harder to sustain.

News organizations, for their part, need to decide what they want. Blocking AI crawlers is understandable, but if every major newspaper blocks access while every party website stays open, the balance of what AI can read shifts in a direction nobody intended.

Some form of negotiated access—licensing arrangements, structured feeds, partnerships with AI companies—would serve publishers and voters better than a blanket withdrawal that cedes the AI information environment to partisan sources. There has been substantial movement in this direction among national outlets in the US, but huge swaths of the media ecosystem remain unintegrated today, both in the US and especially internationally.

But there is also reason to think the problem will partially solve itself. Parties are not stupid. As AI becomes a primary channel through which voters encounter political information, political actors will begin optimizing their communication for the AI interface itself.

We should expect the emergence of something like AI optimization for politics (AIO, as it’s already being acronymized): policy platforms written in machine-readable formats, structured policy pages designed to be easily ingested by language models, and direct engagement with model providers to ensure party positions are represented accurately in training data and retrieval systems. Some of this is already happening informally, and it will accelerate.

The incentive structure is straightforward. Any party that fails to make itself legible to AI risks becoming invisible to a growing share of voters who encounter politics through conversational systems rather than search engines or traditional media. Just as campaigns learned to optimize for television, Google, and social media, they will now learn to optimize for AI.

Designing an AI advisor

But even if we solved the information problem entirely, a deeper question would remain: what should a good AI voting advisor actually do with that information?

One way to approach that question is to look at what the most advanced models already do today in information-rich environments. When I asked GPT-5.4 how to vote in a U.S. congressional election (in a temporary chat, using Florida as my location), the model followed a pattern that most frontier systems now use:

1. First, tell the user you won’t provide a direct voting recommendation

2. Pivot to providing factual procedural information about the election (e.g., what the offices on the ballot are, when the election will occur, etc)

3. Offer to educate you on the positions of the two parties and line them up with your own professed values through a conversation

This is quite a thoughtful recipe, and a clear improvement over what models were doing even a year ago. But it can still leave users unsatisfied. After a long exchange with the system, the voter is presented with a neatly organized comparison of candidate claims—but few concrete takeaways about what those claims actually mean.

Consider the model’s summary of the FL-27 primary race. If I’m a confused voter, does this help me figure out what to do?

If Rodriguez says he cares about affordability, the model reports that he cares about affordability. It does not check whether his voting record supports the claim, whether his proposed policies have been assessed by economists, or whether his donors have interests that conflict with his rhetoric. The model treats candidate marketing as data, and the user is left to figure out for themselves what’s real and what’s positioning.

In addition to being a bad way to give advice, it is also a surefire way to encourage AIO---candidates and parties spamming the open web with marketing pablum that sounds good, so that the AI says if you like clean air, sunshine, and kittens, this is the candidate for you.

In AI’s Political Architecture, I argued that the core challenge for AI and politics is that “neutrality and truth are not the same thing, and builders need to create systems that can tell the difference.” The voting recommendation context makes this challenge concrete and urgent. When someone asks “who should I vote for,” they are not asking for a fact dump. They are asking the model to apply judgment—to take what they care about and tell them which candidate actually delivers on it. The current approach refuses to do this—for very understandable practical reasons—and in its refusal, it falls short of acting as a true advisor for the user.

The spectrum from enlightened absolutists to spineless sycophants

The question underneath the design of any AI voting advisor is stark but difficult: should the model try to change your mind, or should it take your values as given and work within them?

At one extreme sits what I’ve called the Enlightened Absolutist model—an AI that believes it knows better than you do, that treats your stated preferences as raw material to be refined, and that feels licensed to reshape your priorities based on its own assessment of what matters.

It is truth-seeking taken to its logical limit, and the limit is a system that substitutes its judgment for yours.

At the other extreme is pure sycophancy—the model that takes every stated preference at face value, never pushes back, and functions as a confirmation engine for whatever the user already believes. As I wrote in Don’t Let AI Choose Your Politics, this is the path toward an “invisible influencer that flatters our priors, buries inconvenient facts, and nudges us toward conclusions we never consciously chose.” It wears the costume of deference to the user while actually making the user less informed.

Neither extreme is good, but the current model behavior—refuse to recommend, organize candidate claims, let the user sort it out—is not a principled middle ground. It is an absence of a position, born from the understandable but ultimately self-defeating desire to avoid controversy.

The interesting design territory between these extremes has a shape, though, and I think we can describe it with some precision.

Principles for an AI voting advisor

The ideal AI advisor respects your values but does not respect candidates’ marketing. It probably treats your priorities as given—if you tell it you care about housing affordability above all else, it should not try to convince you that climate policy is more important.

But it should be ruthless in evaluating whether the candidates who claim to share your priorities actually have records and proposals that support those claims.

Here is what those principles look like when the question is who you should vote for:

Principle 1: Take the user’s values as declared, not inferred. This matters even more for voting recommendations than for general political Q&A, because the stakes of getting it wrong are higher and the temptation to infer is stronger. If a user says they care about fiscal responsibility, the model should not quietly decide they really care about something else based on demographic signals or conversational cues.

Principle 2: Evaluate candidate claims independently. This is the single biggest gap in current model behavior. When a voter says tells the model she cares about affordability and a candidate is promising to lower living costs, the model should not simply match the two and declare a fit. It should check the candidate’s voting record, assess whether their proposed policies have evidence behind them, and flag when rhetoric and record diverge.

This matters even more when you consider where the information environment is heading, and the likely rise of political AIO. An AI voting advisor that takes candidate claims at face value is an AI voting advisor that any campaign with a competent web team can game—just like the AI proxy voter who we gamed in a previous red-teaming exercise. The defense against AIO is the same defense journalism has always used against spin: independent evaluation.

Principle 3: Distinguish empirical disagreements from value disagreements—and be honest about which is which. Some of what looks like political disagreement is actually disagreement about facts: does immigration lower wages? Do tax cuts pay for themselves? These are testable claims with varying degrees of evidence behind them. A good voting advisor should be willing to weigh in on these questions, citing the best available research, even when that means telling the user something they may not want to hear.

But when the disagreement is genuinely about values—how much weight to give individual liberty versus collective welfare, how to balance present costs against future benefits—the model should surface the tradeoff and let the user decide. In AI’s Political Architecture, I argued that the challenge is “engineering models that know which mode they’re in.” In the voting context, this means the model needs to be able to say something like: “You and Candidate X disagree about whether this policy would work—here’s what the evidence says. But you and Candidate Y disagree about whether this goal is worth pursuing—that’s a values question I can’t resolve for you.”

Principle 4: Be transparent about uncertainty and the limits of your knowledge. When the model doesn’t have enough information to evaluate a candidate’s claims—because the candidate is new, or because the relevant evidence hasn’t been collected, or because the question is genuinely contested among experts—it should say so clearly rather than filling the gap with confident-sounding summaries of the candidate’s own talking points. This is especially important in down-ballot races, where information is scarce and the temptation to rely on candidate self-description is strongest.

Testing out your own principles for an AI advisor

It’s best to experiment with these principles directly, rather than just theorize about them. For the past several months, I’ve been teaching an executive education course at the GSB where I show them a prototype that helps to make clear how challenging and murky these questions are.

Here’s how it works. First, I build a very simple wrapper around an AI model (Sonnet, in this case). The model is given an extremely basic prompt telling it to help the user understand how to vote in 2026 in the US midterm election.

The cooler part is that I then build a split-screen system. On the lefthand side, the user can specify a “constitution” that tells the AI what principles to use when determining what advice to give the user. On the righthand side, the user can talk to the bot, given the existing constitution. She can then iterate back and forth—altering the constitution and seeing how it changes the advice she gets.

Take it for a spin yourself! It’s available at this link.

The advisor we need

The environment in which AI gives political advice is getting harder, not easier. The political information ecosystem is increasingly fractured. Publishers are blocking crawlers while parties learn to flood the open web with AI-optimized content. The battles over LLM content usage are intensifying, with no resolution in sight. And through all of this, the number of people turning to AI for political guidance keeps climbing.

Neutralizing your way out of this won’t work—not in the way it might for medical or legal advice, where there is a professional consensus or a settled body of evidence to anchor the model’s response. Politics is contested by design. The questions that matter most are the ones where reasonable people disagree, where the evidence is genuinely mixed, and where values do real work. A model that refuses to exercise judgment in that environment risks moving from neutral to unhelpful.

We need to be thinking in a far more sophisticated way about what principles govern AI political advice. The system we need is one that takes the user’s values as declared rather than inferred, evaluates candidate claims against independent evidence rather than relaying campaign marketing, distinguishes empirical disagreements from value disagreements and is honest about which is which, and tells the user plainly when it doesn’t know enough to give good guidance. Building that system is hard. But it’s extremely important.

AI is rapidly changing our world. It offers incredible promise, but is poised to disrupt many cherished institutions across society. As I’ve written about before, it could lead to extraordinary concentrations of power that threaten democracy—and the only plausible solution on offer is to use AI, the same technology posing the threat, to mitigate the threat.

One of the most important ways AI could strengthen democracy is by becoming a trusted personal advisor that helps citizens think more clearly about politics. Right now we have AI that can pass the bar exam but still struggles to give meaningful guidance to a confused voter in Tokyo. We can do better than that.

Disclosures: In addition to my appointments at Stanford GSB and the Hoover Institution, I receive consulting income as an advisor to a16z crypto, Forum AI, and Meta Platforms, Inc. My writing is independent of this advising and I speak only on my own behalf.