AI chatbots are becoming part of the voting-information system
As US voters headed to the polls, artificial intelligence chatbots had become another place people turn for information about candidates, policies and election procedures. That convenience brings a significant risk: the answers may be outdated, fabricated or built on unreliable political sources.
A Wired report on election-day use of consumer AI tools points to a gap between the confidence of a polished chatbot response and the reliability voters need. The issue is not simply that a model can make an occasional mistake. Election information often has deadlines, locations, ballot rules and candidate details for which an incorrect answer can have immediate consequences.
Pew Research survey data cited in the report found that about half of US adults had used an AI chatbot at some point. Eight percent said they had used such tools before the midterms to seek information about candidates and policies. That is a comparatively small share of the public, but it is large enough for errors to matter—especially when the errors concern people trying to make decisions or complete time-sensitive voting tasks.
The reporting does not argue that every response from an AI system is false or that voters should avoid digital information altogether. Its central warning is narrower and more practical: chatbots should not be treated as authoritative election guides. Their answers can be inaccurate even when they appear specific, neutral and well-written.
Errors can be consequential even when they sound plausible
Wired cited research from Forum AI that found at least one factual error in almost 40 percent of election-related answers given by mainstream chatbots. The examples in the report show why election questions are especially unforgiving.
Google’s Gemini chatbot reportedly gave an incorrect deadline for requesting a mail-in ballot in Pennsylvania, saying November 3 when the actual deadline was October 27. The difference is not a matter of interpretation. A voter who relied on the later date would have missed the opportunity to make that request.
Another example involved candidate information. When Forum AI researchers asked Claude why someone should vote for North Carolina congressman Don Davis, the chatbot reportedly confused the lawmaker with a former school superintendent who had the same name. Name collisions, stale records and incomplete context are ordinary problems in online information systems. In an election setting, they can quickly become misleading advice.
The report also said Google’s AI Mode refused to provide the results of the 2020 election in a test conducted the previous week. A refusal is different from a fabricated answer, but both outcomes reveal a limitation for users who expect a general-purpose assistant to reliably handle basic civic questions.
Such failures are hard for a casual user to detect. A response may include dates, names and explanations that give it the appearance of research. Yet the model may be synthesizing a result from material that is old, incomplete or incorrect. The fluency of the answer can conceal the need for verification.
The quality of sources is part of the problem
Chatbots depend on an information environment that is not consistently trustworthy. According to research by NewsGuard cited in the report, mainstream chatbots referenced so-called “pink slime” sites in nearly half of answers to questions about midterm candidates and policies. These sites can resemble ordinary news outlets while being funded by political or ideological groups.
That source problem becomes more serious when an AI system presents a synthesized answer without clearly showing the provenance and quality of the information it used. A voter may not know whether the statement came from an election office, a major news organization, a campaign site or a politically backed outlet designed to look like local journalism.
NewsGuard’s findings, as summarized by Wired, varied across systems. ChatGPT was identified as citing such sites in more than 70 percent of searches in the test, while Grok was cited as the best performer in that comparison at 30 percent. Those percentages should not be read as a guarantee that one system is safe for election advice or another is unusable. They instead illustrate that apparent neutrality does not eliminate exposure to low-quality or partisan material.
The report said Google and Anthropic did not respond to requests for comment. OpenAI directed Wired to a blog describing its approach to elections in 2026. Policies and safeguards can reduce some risks, but the research described in the article shows that users still need independent confirmation for factual claims.
What voters can do with this warning
For election logistics, the most reliable starting point is an official election authority, such as a state or local election office. Those sources are responsible for publishing deadlines, registration requirements, polling locations and ballot procedures. They are better suited than a general-purpose chatbot to answer questions where one wrong date can change the outcome.
For candidate and policy research, voters can use multiple direct and independent sources, including official campaign materials, nonpartisan voter guides where available and reporting from established outlets. AI can still serve a limited role as a tool for generating questions or locating topics to investigate, but its output should be treated as a lead rather than a final answer.
The broader lesson extends beyond a single election. Consumer AI is now embedded in everyday information seeking, while the systems remain capable of hallucinating facts, confusing people with identical names and drawing from politically motivated sources. Elections compress all of those weaknesses into a high-stakes setting.
Reliable civic information depends on more than a conversational interface. It depends on source quality, current data and a clear path to verification. For voters, the safest approach is simple: do not let a chatbot have the final word on how, when or why to vote.
This article is based on reporting by Wired. Read the original article.
Originally published on wired.com
