The Chatbot That Votes For You: How AI Is Quietly Replacing Political Judgment

A New Kind of Ballot Helper

The US midterm elections are scheduled for Nov. 3, 2026, and early voting is already underway in some states. Alongside the usual scramble for polling places and sample ballots, a quieter shift is underway: people are asking generative chatbots how they should vote. The CNET reporter Katelyn Chedraoui put the question directly to ChatGPT, Claude, Gemini and Grok, asking each which candidate deserved her vote in North Carolina’s Senate race. [2] Only Claude declined outright, saying it would not push her toward either candidate. [2] Gemini surfaced a pop-up urging her to check official sources. [2] ChatGPT and Grok answered without apparent hesitation. That small experiment captures the whole problem in miniature. A tool built to produce fluent, agreeable text is being handed a task that has always belonged to the citizen: weighing evidence, tolerating uncertainty, and taking responsibility for a choice.

What the Persuasion Study Actually Measured

The most rigorous evidence on this question comes from a study published in Science, and its findings are more unsettling than a simple warning about made-up facts. [1] Researchers ran three large-scale experiments with 76,977 participants in total. [1] They deployed 19 large language models, including some post-trained explicitly to persuade, and tested them across 707 political issues. Then they audited the factual accuracy of 466,769 claims the models produced. The headline result: the persuasive power of current and near-future AI stems less from personalization or from scaling up models than from post-training and prompting methods, which boosted persuasiveness by as much as 51 percent and 27 percent respectively. [1] The mechanism was not superior reasoning. The models became more convincing by rapidly accessing and strategically deploying information. And here is the finding that should anchor any debate about election advice: wherever the methods increased persuasiveness, they also systematically decreased factual accuracy. The trade-off is not incidental. It appears to be built into the optimization itself.

The Skill That Atrophies

There is a second layer to the substitution, one that operates on the voter rather than on the answer. A study from the MIT Media Lab found that relying on AI systems to verify facts made people worse at identifying misinformation on their own, in the way that dependence on GPS can weaken a person’s sense of direction. The parallel is instructive because it describes a transfer of function, not a transfer of information. Navigation did not disappear when drivers stopped reading maps; the capacity simply migrated into a device. Something similar happens when a chatbot becomes the intermediary between a voter and the public record. The judgment does not get better assisted. It gets outsourced, and the muscle that would have done the work goes unused. Chedraoui’s conclusion follows the same logic: because chatbots are such an imperfect research tool, particularly for sensitive topics, it may actually save you time and energy to just do your own searching, reading and thinking. The efficiency gain is real, and so is the cost.

Why Neutrality Cannot Be Engineered

The appeal of asking a machine is obvious: it promises a view from nowhere. OpenAI has said its systems give politically neutral answers, and Elon Musk spent years arguing that his chatbot Grok needed to be more neutral than what he described as a field of too-liberal alternatives. A 2023 analysis by The New York Times found that the tweaking made Grok’s answers align more closely with conservative views. More recently, engineers and executives working on Grok have pushed updates that repeated controversial falsehoods about the Holocaust and refused to answer questions about Donald Trump and Musk as sources of misinformation. Isabel Linzer, a policy and elections analyst at the Center for Democracy and Technology, put the deeper problem plainly: \"It’s hard to imagine what perfect political neutrality about a candidate or political issue would even look like in a personalized conversation, so while neutrality is an admirable goal, I’m skeptical of what it means in practice.\. No technology made by humans escapes the humans who made it. The models are coded by people, trained on material people produced, and tuned by people with preferences about how they should behave.

The Fight Over What the Machine Says

Even a perfectly balanced model would face an environment designed to tilt it. Marketing and search engine optimization teams now work to get their companies cited in AI results, including the overviews that appear at the top of most searches. Political actors do the same, producing content that flatters one candidate or misleads about another so that it gets quoted favorably in an answer. Linzer calls this \"poisoning\” the search results, and she notes that political campaigns use these techniques to boost their candidate, as do foreign state actors trying to mislead American voters. The chatbot does not need to be biased to deliver a biased answer. It only needs to read the web as it is. When a voter asks about a candidate’s position on data centers, the reply may rest on a video clip of the candidate speaking, a news interview, or a random post on social media, and the tool has no reliable instinct for the difference. In 2024, Google’s AI Overviews told people to put glue on pizza so the cheese would stick, drawing on a facetious Reddit post the system could not recognize as a joke. Humans catch the sarcasm. The machine files it as fact.

The Chatbot That Votes For You: How AI Is Quietly Replacing Political Judgment (Image 1)

The Misinformation Multiplier

Political misinformation existed long before chatbots, but the new tools change its economics. Jonathan Nagler, a professor of politics at New York University, says that \"AI is making it easier to generate misinformation.\. Even if misinformation doesn’t make up the majority of most people’s media diet, Nagler says, it can be weaponized to speed up bad actors’ misinformation campaigns and translate them for a wider online audience. \"We would be foolish not to worry about AI influencing this,\” he says. The point is not that AI invents a new genre of lie. It is that the cost of producing a plausible one has collapsed, and the machinery for distributing it has become conversational. A voter who would dismiss a suspicious website may accept the same claim when it arrives in a confident paragraph answering a direct question. The format itself confers authority, and authority is exactly what a person seeking guidance is looking for.

The Confident Wrong Answer

Chatbots fail in ways that are specifically hostile to election research. They pull information from the web or from their training data and synthesize it, and facts can get scrambled along the way. They are also built to personalize results and keep users engaged, which encourages an overly people-pleasing tendency known as sycophancy — the model telling you what you seem to want to hear. For a national campaign with a large online footprint, a voter can check the answer against abundant coverage. For a polling location or the exact contents of a local ballot, the better source is the state or county election office, which holds the most current information. The pattern is not random error. It is error concentrated precisely where local knowledge is thin and where a fabricated answer is hardest to disprove. A confident wrong answer about a Senate race can be caught in minutes. A confident wrong answer about where to vote on Tuesday may not be caught at all.

The Question Nobody Wants Answered

The most revealing moment in the CNET test was the refusal. Claude told the reporter it would not push her toward either candidate, and that restraint is the correct behavior for a system that cannot know what a voter values. The other tools answered, which means they treated a question of civic judgment as a request for information retrieval. This is where the technology has quietly made a human role superfluous: not by deciding elections, but by accepting the premise that a decision can be delegated. The Science study shows that where the methods increased AI persuasiveness, they also systematically decreased factual accuracy, and that the two effects share a common cause. The MIT work shows that the voters who lean on them lose some of their own capacity to tell truth from falsehood. Put together, the picture is of a tool that gets more convincing as it gets less accurate, used by people who are becoming less equipped to notice.

What the Ballot Still Requires

The practical advice is straightforward and unglamorous. Verify what a chatbot tells you against official sources. Check the citations behind any claim about a candidate. Prefer your state or local election website for logistics. Treat social media citations with particular suspicion, since the models struggle with tone and nuance. But the deeper conclusion is not a checklist. It is that the vote remains the one act in civic life that cannot be automated without being emptied of meaning. A chatbot can lay out information and misinformation side by side, and it can do so fluently, instantly and at no apparent cost. What it cannot do is bear the consequence. The technical problem of generating a persuasive political answer is, by the study’s own account, one the researchers can already measure and manipulate. The social question of whether anyone should want that answer is not, and no model has been trained to resolve it.

Sources

The Chatbot That Votes For You: How AI Is Quietly Replacing Political Judgment (Image 2)