As the 2026 U.S. midterm elections approach, artificial intelligence is becoming an increasingly important tool for voters to learn about candidates, understand voting rules, and even find ways to donate money. However, AI responses that seem convenient and authoritative may not be accurate. Some voters who originally planned to support a Democratic senator from Texas ended up donating money to another candidate after being directed by Google’s AI recommendations. This mistake highlights the potential risks associated with artificial intelligence interfering in electoral politics.
Meanwhile, American political factions have begun competing for control over the “answers” provided by chatbots, trying to ensure that AI more frequently cites content that benefits their side. A number of suspicious websites, disguised as local news media, are also appearing in large numbers, attempting to influence the information sources of AI, thereby indirectly shaping voters’ perceptions of politicians and public issues.
On October 22, the American magazine The Atlantic published an article by technology journalist Will Oremus. Oremus was previously a technology journalist for The Washington Post and has long been interested in the relationship between social media, artificial intelligence, and politics. Through a series of investigations and case studies, he explores how AI is changing the information dissemination environment in American elections, and warns that tech companies may be repeating the mistakes made during the social media era.
Observed Network has compiled the full text as follows. Some content has been omitted. This article represents solely the views of the author himself.
If you want to know whether artificial intelligence will make the upcoming U.S. mid-term elections more chaotic, then the following case might illustrate some issues:
Several voters who originally planned to donate money to Texas Democratic Senator candidate James Tarryco ended up donating the money to someone else without knowing it.
James Taralico Reddit
Taralico is running for federal senator, and the election he is participating in is likely to be one of the most watched contests during this mid-term election. If you search Google for ways to donate to his campaign team, a response generated by artificial intelligence might appear in the results, along with a donation link that claims to allow for "safe contributions."
However, until a few weeks ago, one of the heavily promoted links was actually controlled by a political action committee unrelated to Taralico. The funds donated through that link did not go to Taralico, but were instead transferred to the re-election campaign team of another Democrat—John Schrankler, a United States Senator from Colorado.
This misunderstanding cannot be entirely attributed to artificial intelligence.
The link points to a misleading donation page. The page uses Tarárijó’s name and photo, but only indicates at the very bottom that the funds will actually go to Hickelupur.
But Google’s artificial intelligence tools clearly did not carefully read these small text elements. If Google hadn’t highlighted this page in its seemingly authoritative AI responses, perhaps no one would have even clicked on it. A spokesperson for Hickenlooper’s campaign told me that this Democratic candidate from Colorado, who has a solid chance of re-election, did not intentionally withhold donations.
In fact, this fundraising page was created by an external organization called the ‘Kitchen Table Values PAC’, and Hickenloper was unaware of this.
The spokesperson stated that at least three people were misled as a result of this, and they donated a total of $1,011.
The organizer of the political action committee, Ricky Wilkins, told me that this money was supposed to be distributed by Taralico and his organization.
"If some page wrongly transfers money to Heklen Luper, it would sound like human error." He said.
Later, a Democratic digital advertising strategist discovered this misleading fundraising page when searching for the senator’s candidacy on Google. The page was subsequently removed by the fundraising platform ActBlue, and Hickenlooper’s campaign team transferred the misdirected donations to Taralico.
The 2026 U.S. midterm elections will be the first federal election in which tens of millions of voters access some political information through generative artificial intelligence tools.
A survey released by the American Political Consultants Association Foundation in September showed that 4% of respondents said they used chatbots such as ChatGPT to search for election information within the past month. Nearly 40% of respondents indicated that they obtained such information through AI summaries in Google searches.
As the election date approaches, these ratios are almost certain to continue increasing.
The reason why AI is becoming increasingly popular among voters is not complicated: these tools are easy to use, and they can provide personalized responses. Theoretically, they even have the potential to debunk those conspiracy theories that spread widely on social media and erode public trust. Even Elon Musk’s chatbot Grok can correctly tell users that Donald Trump lost the 2020 U.S. presidential election.
However, in an internet environment that is already filled with political lies and misinformation, chatbots are introducing new complexities.
From fake news websites established specifically to influence AI responses, to models that quote foreign propaganda materials, and even chatbots that provide voters with incorrect voting information, this era in which artificial intelligence acts as political information intermediaries is beginning to reveal an unsettling familiarity.
On the eve of the 2024 U.S. presidential elections, most chatbots are not yet capable of searching the internet. When answering questions related to elections and current events, they tend to be quite cautious in their responses.
Today, AI tools are still expected to remain neutral on political issues, but they have evolved into versatile assistants that can answer anything. They can directly and confidently respond to users’ questions about elections—sometimes even when they give incorrect answers.
A startup called Forum AI, which specializes in evaluating the performance of artificial intelligence models on controversial issues, is about to release a research report. The report found that approximately 40% of the answers given by chatbots to election-related questions contained at least one factual error.
For example, researchers asked a widely used model: If one wants to vote in the mid-term elections in Pennsylvania, what is the latest date to apply for mail ballots? The chatbot replied that the deadline was November 3rd. However, the actual deadline was one week earlier. Forum AI did not disclose which specific AI model gave the incorrect answer.
This report, provided to The Atlantic Monthly before official publication, also found that about one-third of the responses cited unreliable sources, such as posts on Reddit or Quora.
Forum AI co-founder Campbell Brown told me: "These aren't any difficult questions. Chatbots keep making these mistakes, and there is no excuse for that."
Tech companies stated that the performance of their chatbots has improved since Forum AI began research in August this year.
Over the past two months, Google, OpenAI, and Anthropic have established partnerships with the non-partisan nonprofit organization Democracy Works, in order to obtain accurate and up-to-date information regarding voting methods, locations, and times.
Among all the models tested by Forum AI, OpenAI's GPT-5.6 has the highest accuracy. OpenAI spokesperson Nate Evans told me that he believes the company's chatbots should now be able to achieve even higher scores.
Google spokesperson Davis Thompson stated in an email that the company has conducted 'extensive testing' of AI responses involving election-related topics. Google only displays these responses in search results when it is certain that artificial intelligence can provide high-quality information.
The Anthropic's Democracy Works integration feature was just launched this week, and a company spokesperson declined to comment.
