Should an AI tool steer you toward one voting choice over another, even if every fact it gives you checks out?

This explores what can go wrong when voters ask AI tools for help deciding how to vote. The main lesson from the corpus is that factual accuracy has largely stopped being the main problem. What matters more is where the tools send people and which options they put in front of them. By early 2026, States United found that ChatGPT and Google's AI answered voter questions with no verifiable factual errors Do AI chatbots give voters accurate election information?. The same audit found two other gaps. The chatbots pointed voters to official state election websites less than half the time. They also gave incomplete information about candidates and relied on sources anyone can edit, such as Wikipedia. So an answer can be correct and still fail to connect the voter to an authoritative source.

The more troubling finding is about steering. The Dutch Data Protection Authority tested general-purpose chatbots as voting-advice tools. They recommended the same two parties in more than half of cases, even when the test user's stated positions matched a different party Do chatbots steer Dutch voters toward the same parties?. The two parties sit at opposite ends of Dutch politics: the right-wing PVV came up about 30% of the time and the left-wing GroenLinks–PvdA about 25%. Centrist parties appeared in under 2% of recommendations. The regulator called this a 'vacuum cleaner effect' that pulls voters toward the poles Do AI chatbots systematically bias voters toward extreme parties?. They traced the cause to the data. The chatbots drew on unstructured internet text rather than structured party-position data, so parties that dominate online discussion crowded out the ones that don't. A tool that presents itself as a neutral matching service can quietly narrow political choice without stating a single false fact.

Several notes that aren't about elections help explain why this is hard to fix. AI output changes with prompt wording, random sampling and context Why does AI output change with every prompt and context?. That means two voters asking nearly the same question can get different advice, and an audit of a fixed set of prompts can't guarantee what any one voter will see. Separate research finds that assistants have no internal record of what they don't know about the user. That gap feeds sycophancy (telling people what they want to hear) and confident guessing. Adding an explicit list of unknowns to the prompt cut those failures roughly in half or more Do language models know what they don't know about users?. A voting assistant that doesn't track what it hasn't learned about your priorities will fill the gaps with whatever its training data makes most prominent.

The corpus also suggests better designs. 'Learning to guide' research argues that AI should point out the parts of a decision worth a closer look rather than hand over a verdict. This avoids anchoring, where people latch onto the machine's answer, and keeps responsibility with the person deciding Can AI guidance reduce anchoring bias better than AI decisions?. For voting, that would mean showing where parties actually differ on your issues instead of naming a winner. Chatbots can also improve political information. Ten-minute chats with an AI representing the other party corrected partisan misperceptions and made people feel warmer toward the other side. The effect came from correcting false beliefs, not from persuasion, though most of it faded within a week Can AI chatbots reduce partisan misperceptions and warm cross-party feelings?.

The corpus has limits here. The direct evidence on AI and voting comes mainly from two audits, one American and one Dutch. Nothing in the corpus measures whether chatbot steering actually changed anyone's vote. The open question is not whether these tools get facts right. It is who decides which facts and parties a voter gets to see.