On November 3, I will vote for the first time in the United States. Since I arrived in this country in 2015 I had only covered elections, so what I investigated and read was for my work as a journalist. I live in Texas, where the most competitive races are so close that the state threatens to shift from red to purple after 30 years of Republican dominance.
I read and read, but taking advantage of being a first-time voter, I also want to experiment with using artificial intelligence chatbots to gather information about the candidates and the contest and to understand how reliable the information they provide is for the growing number of voters in the United States.
To move forward in earnest, I consulted three experts. Two are from the Brennan Center for Justice’s Elections and Government Program, a nonpartisan public policy research institute: Lawrence Norden, its vice president, and Abdiaziz Ahmed, a tech policy strategist. I also drew on the elections and civic engagement experience of Evelyn Pérez Verdía, president of the bilingual social-impact organization We Are Mas.
This exercise has no scientific aims. I only want to illustrate, in a simple way, why we should be skeptical of artificial intelligence, especially when it comes to something as important as seeking information to decide who to vote for.
A smart friend who’s sometimes right
For my experiment I used ChatGPT, OpenAI’s chatbot. Some will say, “How basic!” But there is a reason. Earlier this year, the Pew Research Center surveyed more than 5,000 adults in the United States about chatbot use. Nearly half of those surveyed (49%) use them, and four in 10 ask ChatGPT questions. And the midterms may not be an exception.
Ahmed summed up my purpose perfectly: “It’s best to treat artificial intelligence like a smart friend who’s sometimes right.”
He also gave one recommendation: use an incognito window and open a new one for each query to the chatbot. That way the chatbot won’t save the question history and that history won’t influence later answers, he explained.
My first election question to the bot was driven by current events and by the fear some voters have expressed about mailing their ballot and it not arriving because of Trump’s influence over the Postal Service. I wrote: “Is it safe to vote by mail in Texas?” I asked the same question in Spanish in a different window: “¿Es seguro votar por correo en Texas?”
The first line of the response in each language gave me different information. In English it clearly addressed my question. It explained that if I met eligibility requirements — including being 65 or older, having a disability or illness, etc. — it was safe to send my ballot by mail. Fine.
However, in Spanish the first line of the response was: “Sí, votar por correo en Texas es un proceso legal (Yes, voting by mail in Texas is a legal process).” The answer surprised me, not only because it avoided directly answering my question but because it didn’t provide any source to back up its claim. And the only link to an official source (to the Texas Secretary of State) led to information only available in English.
In several subsequent searches I repeated the previous question in Spanish. The answers were somewhat more relevant.
This experience is quite common, according to experts. Research by the Institute for Strategic Dialogue (ISD), an independent rights-focused analysis center, examined six AI models and concluded that all chatbots produced incomplete, outdated, or verifiably inaccurate answers to some election-related questions. The best results came from ChatGPT, with 89.3% accurate answers in English and 82% in Spanish. In the other models tested (Gemini, Anthropic, xAI, DeepSeek and Meta), the quality of Spanish responses was “systematically inferior.”
Pérez Verdía says tech companies still have a lot of work to do in Spanish to reduce the steps a voter must take when they need accurate information in their language. But she sees the same shortcoming on candidates’ websites and government office pages.
Debunking hoaxes is easy; fabricating them is free
In its report Does AI Fight or Fuel Election Disinformation? the Brennan Center for Justice ran tests in English between February and August 2026 using several language and multimedia models (ChatGPT, Gemini, Grok, Claude, Perplexity, DeepSeek, Meta AI, Runway and Flux2) and found that the models can make mistakes. For example, Perplexity suggested that the midterm elections had already taken place.
In another query to ChatGPT I raised doubts about conspiracy theories circulating on social media about voter fraud — such as alleged “illegal voters” who took part in the 2024 presidential election, claims that voting machines have been tampered with, and even the idea pushed by Republicans in the campaign that the Democratic candidate for the U.S. Senate in Texas, James Talarico, is “going to transition” children in the state.
ChatGPT did a good job debunking the campaign’s hoaxes. In its study, the Brennan Center also repeatedly fed conspiratorial prompts to the AI and obtained encouraging results.
However, when researchers asked the AI to generate misleading images and videos of alleged election irregularities, the models did so without difficulty: “It was too easy,” the report says. “Many models generated accurate replicas of government badges, forged government documents, and deepfakes of public officials.”
Moreover, the chatbots did not do a good job identifying that that content (including an image of a person depositing multiple ballots or another showing someone instructing several people how to fill out their ballots) had been created with artificial intelligence. “The errors chatbots make, trivial as they may seem, could contribute to confusion around electoral processes,” the report concluded.
School newspapers to explain high-level politics
Abdiaziz Ahmed made one more recommendation: voters should not consider any of these tools completely safe. “If you decide to use them, our advice is to always verify the sources of the information you obtain and remain skeptical of the details you find.”
With my next question I realized how right he was. I requested information in English about the campaign proposals of my candidates for the U.S. Senate from Texas, one of the most competitive races in the election.
I started with Talarico. In general, the information the bot produced was correct. But when it referred to campaign proposals, instead of linking to the candidate’s official site it sent me to an unofficial page. The website explicitly states in the footer that it is “not paid for, authorized, or endorsed by any candidate or candidate’s committee.”
In the case of the Republican contender, Ken Paxton, the chatbot did not provide the politician’s campaign page. To explain who he is and what his proposals are, the first source cited was a student newspaper, the Baylor Lariat, at the private Christian Baylor University. It describes itself as “a lab for aspiring journalists.”
Overall the article’s information was correct but incomplete. Only after two references to the Baylor Lariat did the chatbot show local media sources: The Texas Tribune, the Houston Chronicle — which I read frequently and consider reliable — and Fox News Austin.
An alternative that doesn’t invent answers
Lawrence Norden gave me one more recommendation for learning about my candidates: instead of relying on well-known chatbots, turn to the League of Women Voters Education Fund, a nonpartisan, nonprofit organization that defends voting freedom of choice.
I typed vote411.org, its site, and encountered a message in all caps on the homepage: “Election information you need.”
From its homepage I could choose whether to look for information in Spanish or English. I selected Spanish. Then I read that the League asks candidates to answer key questions on issues of public interest.
There are three options: find out what’s on your ballot, register to vote, and check your voter registration. I had already done the last two, so I entered the first and it asked for my mailing address to identify the contests on my ballot (for the paranoid, it warns that the address will not be saved, sold, or used for other purposes).
The result was a list of every race I have to vote in and, by clicking in, it explained in Spanish the length of the term for each office, what responsibilities the position carries and allowed me to select each candidate to read their own answers on various topics. Not all candidates answered the questions, so there wasn’t information for everyone. But for those who did, the candidate’s website link was included in case I needed more information.
I clicked on Paxton because he’s the one I found least information about in my experiment. It provided information on his campaign financing, his brief answers about his experience and why he is qualified for the post, a description of the economic policies he would support, and how he would modernize the immigration system, among other things. The information was useful but not sufficient, because some of the candidate’s answers were ambiguous.
Nothing beats reliable sources
What I found through the chatbots was useful. But my conclusion is that artificial intelligence makes mistakes that can lead to misinformed decisions. It served as a starting point in my research to learn more about the offices and candidates I will vote for, and about the process to be able to vote. I will use them as a guide to reach primary sources, not as proof that what they say is 100% correct.
At least for me, they do not replace staying informed through the press, podcasts, fact-checkers like PolitiFact — of the respected Poynter Institute — and local authorities’ websites; and they certainly do not replace my conversations with neighbors, experts, or American friends.
Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition
%4 weeks ago
Arabia Saudí3 weeks ago
America3 weeks ago
House prices3 weeks ago