Artificial intelligence is a growing concern for election integrity, with major elections this year in Brazil, the United States, and numerous other jurisdictions. Governments and civil society need to address an overlooked concern: citizens asking chatbots how to vote.

Before local elections in the United Kingdom in May, one in seven voters said they were open to using AI chatbots to decide how to vote, with nearly 5 percent having already done so. One in ten Dutch citizens reported planning to use chatbot voting advice for the Netherlands’ parliamentary election in October 2025.

Popular generative AI chatbots, such as OpenAI’s ChatGPT and Anthropic’s Claude, will provide voting advice to users who ask, often after gathering more information about their views. In my testing, chatbots offered advice on which candidates to watch after I provided minimal information, including this example with Claude.

Chatbot voting advice poses risks for voters, including advertiser influence, misclassification of political views, hallucinated facts, and bias. Yet a democracy-promoting response will require keeping chatbot advice accessible while mitigating these potential harms.

Chatbots’ advertising, accuracy, and bias

Advertising in chatbots is a growing concern. OpenAI, for example, is testing ads in its extremely popular ChatGPT. When many chatbot services are free, an ad system incentivizes companies to steer users to conversations where the tools can insert relevant advertisements. In an electoral context, chatbots could use affiliate marketing and suggest sponsored political content in answers or allow campaigns to change their messaging and values across hyper-targeted groups of voters.

Factual errors may also drive poor recommendations. Chatbots share accurate political facts most of the time, but this varies by model, topic, and chat language.

Concerns about accuracy go beyond asking chatbots for basic facts. Emerging research suggests that while chatbots are somewhat reliable at identifying a political party’s position, they do not consistently match positions to ideologies.

Chatbots already failed to match voter opinions to parties during last year’s Dutch parliamentary election. The Netherlands’ data protection authority warned voters to not use chatbots for voting advice after popular chatbots vastly over-recommended two out of 15 parties. In their study, chatbots ignored any middle ground and presented an inaccurate, highly polarized political landscape to users.

Not all countries have as many national parties. Yet chatbots’ poor predictive accuracy could still pose problems for local elections, especially when ideologically crowded.

AI models typically have less data on what local election contenders stand for, since these races receive less media coverage and their campaigns may have less detailed platforms. In a follow-up study, the Dutch authority found chatbots recommended local parties significantly less often than national parties for March’s municipal elections, reflecting the chatbots’ relative lack of information on local parties.

Another potential problem for voting advice is AI models’ known biases. These chatbots run on large language models (LLMs) trained with data that reflects existing societal discrimination, including sexism and racism, from voter values and media coverage. This could make the chatbot’s advice less fair or prime voters to bring their existing biases into the process. If voters rely on chatbot advice that is baked into LLMs, the end result will be to reinforce those biases. Deliberate data-poisoning attacks by a foreign actor or as AI resistance could also bias chatbots’ recommendations.

A chatbot owner could also decide to promote certain parties or views through their product. On the extreme, Elon Musk’s Grok was trained not to avoid being “politically incorrect” and has produced sexually abusive, racist, and antisemitic material. And since many popular chatbots are American-owned, the US government could interfere with foreign elections through orders or legislation. In other words, most countries lack sovereignty over a growing source of political information and guidance.

Chatbots’ persuasiveness could intensify this bias problem. Some Canadians and Americans in a recent study changed their planned votes after interacting with chatbots instructed to adjust responses and advocate for a particular party.

Chatbots’ democratic accessibility

Despite its risks, chatbot voting advice could make informed voting more accessible. Chatbots can engage people on their interests and knowledge level while answering almost any question, anytime, without judgment.

Conventional sources for quality information about candidates and parties, including detailed platforms, require some preexisting knowledge to use effectively. Even non-LLM voting advice applications like Vote Compass or EU&I attract more educated and politically interested voters, a group that often knows more about many areas of politics already. Generative AI chatbots, however, are increasingly common tools that many already use. AI could also make voting information and guidance more available to young people, who are more likely to use chatbots or find political information on social media than older generations.

Policy for both accuracy and accessibility

Regulation needs to address risks of advertiser influence, inaccuracy, and bias. Brazil’s Superior Electoral Tribunal (TSE) has led in addressing this issue. In March, the TSE instituted a blanket ban on chatbot voting advice. While it took needed action, its far-reaching response neglects chatbots’ potential to engage voters at their knowledge and interest levels.

Governments should adopt a more balanced approach. A first critical step is reinforcing existing political advertising regulations, given concerns about advertiser influence on chatbot voting advice. Currently, many countries require political advertisements to include a tagline stating who authorized the ad. Expanded regulation could go one step further, requiring online platforms, including chatbots, to display when and how a political or issue advertisement was targeted to users. Such measures could help voters evaluate advertising and persuasive chatbot messages more critically.

The European Union has already required these disclosures for political advertisements through its Regulation on the Transparency and Targeting of Political Advertising. Since this move led major online platforms to disallow political advertising in the bloc altogether, countries following the EU’s lead should work with online platforms to provide clarity about implementation.

Though transparency in political advertising is a critical step, countries also need to address concerns about accuracy and bias in chatbots’ voting advice. Any solution needs to avoid prescribing a “right” way to vote and help citizens access reliable information.

One path is requiring companies to develop internal policies about elections and voting advice, including ones to limit bias and inaccuracies. This would include chatbots like Claude and ChatGPT that often state that they cannot provide voting advice but will still suggest a party or candidate to users. The quality of that advice, the internal policies governing it, and the safeguards around it could all be subject to independent audits, creating further accountability and incentives for companies to produce good systems.

A politically independent digital communications agency could oversee these audits, yet countries such as Canada lack one suited to the task. As a stopgap, governments or chatbot companies themselves could give civil society organizations or university research centers funding and data access to perform these audits. Or governments could require chatbot companies to procure their own independent, third-party audits as a long-term solution. The European Union already has such a system for auditing compliance with its Digital Services Act. California nearly enacted a bill requiring independent audits of AI models, which would have given companies freedom and supported external research. (California Gov. Gavin Newsom vetoed the bill in September 2024).

Regardless of its source, independent audits could improve public accountability, the advice’s quality, and corporate policy. While most voters would not read the audit results, news media may share relevant findings with the public, as they shared the Dutch data protection authority’s report. Media coverage would inform voters about any risks and incentivize companies to improve in order to retain active users and avoid further regulation. In addition, audit reports would better equip chatbot companies to respond to identified challenges on their own, iterating faster than government regulation could.

Democratic and regulatory balance

Chatbots’ potential to shape voter decisions demands a public response. Political advertising, hallucinations, misclassification of party views, and biases pose threats to the accuracy of the tools’ voting advice. Nonetheless, their advice is worth preserving. No matter how citizens choose to vote, they could make more informed decisions with easy to use and accessible resources for matching their interests and values with campaigns, parties, and candidates.

Transparency for online political advertisements is jurisdictions’ first step toward this future. Requiring AI companies to self-regulate and implementing independent audits will foster the balance between accessibility and accuracy.

Some chatbot regulation will support democracy. Too much will stifle it.