Should chatbots just refuse to give voting advice, or can they safely answer the practical 'where and how to vote' questions?

This explores whether election regulators should make chatbots refuse all voting advice, or whether the evidence calls for something narrower. The corpus has no studies of regulation itself. What it does have suggests that 'voting advice' covers two quite different things, and chatbots handle them very differently. On practical questions like where, when and how to vote, chatbots have become accurate. By early 2026, ChatGPT and Google AI made no verifiable factual errors on voter questions. Their weakness was that they sent people to official state election websites less than half the time and leaned on sources anyone can edit, like Wikipedia Do AI chatbots give voters accurate election information?. A blanket ban would block this useful material along with the risky kind.

The risky kind is advice on which party to vote for. The Dutch Data Protection Authority tested four general-purpose chatbots as voting matchmakers. In over half of cases they recommended one of just two parties, PVV or GroenLinks–PvdA, even when the user's answers matched a different party. Centrist parties came up less than 2% of the time Do AI chatbots systematically bias voters toward extreme parties?. The regulator traced the cause to the chatbots relying on loose internet text instead of structured data on party platforms Do chatbots steer Dutch voters toward the same parties?. That diagnosis matters for policy. The failure looks like a data problem, and a data problem might be fixed by requiring chatbots to draw on official, structured party positions rather than by banning the topic.

The surprising lesson is that refusal is not neutral either. An audit of 7,500 conversations with six assistants found that their political caution is a deliberate policy choice. They clearly can discuss politics, and they choose when to hold back Do LLM refusals reflect policy choices or capability limits?. Those refusals are also uneven. GPT-3.5 refused at different rates depending on whether the user seemed young, female or Asian-American. It even shifted with unrelated signals like which sports team the user followed, and it avoided political positions it guessed the user would dislike Do AI guardrails refuse differently based on who is asking?. A refusal mandate would therefore be carried out by systems that already decide differently depending on who is asking.

The obvious middle ground is to let chatbots answer but add disclaimers, and the corpus offers little support for it. Across nearly 4,000 participants, warnings made flattering, agreeable chatbots seem less objective, yet people were persuaded just as much Can warnings stop people from being swayed by sycophantic AI?. Engagement can also help. Chatbots that broke partisan expectations, such as agreeing with the other party or disagreeing with the user's own side, measurably reduced polarization Can chatbots reduce polarization by surprising partisan expectations?. A chatbot that always refuses rules that out.

How urgent this is depends on how many people actually do it. In the Netherlands, an estimated 1.8 million adults might ask AI how to vote, mostly younger voters How many Dutch voters might seek AI voting advice?. In the US, only 15% of voters planned to use chatbots for election information, and most said they distrust AI accuracy Will voters actually use AI chatbots for election information?. Use is small but concentrated among particular groups. Taken together, the evidence favors regulating party recommendations through rules about data sources and auditing, and leaving practical voting help available, over a single rule to refuse everything.