When AI Answers: Texas

Digital Democracy Institute of the Americas (DDIA)

Source published: Site published:

Original article

When AI Answers: Texas
-Nov 17, 2026
Share

By: Rafael Olavarría, Research Consultant, DDIA

As the races for seats in the U.S. Senate and House in Texas became more competitive, voters used Google Search to ask questions about Democratic candidates James Talarico and Bobby Pulido and Republican candidates Ken Paxton and Monica De La Cruz. Talarico and Pulido won their races on November 3, 2026.

Voters may also use artificial intelligence tools to learn about politics. So, in October, the Digital Democracy Institute of the Americas (DDIA) asked three chatbots—OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini—the questions that people most often asked on Google Trends.

These were some of the main findings about Texas. More detailed results appear below. The full results from the When AI Answers series are available in the full report. The results for Michigan and Florida are available here.

  • Claude, and Gemini when answering in Spanish, repeated language from political campaigns and political groups as if it were their own wording.

  • Gemini often used social media posts as sources. Some of these posts came from the candidates' political opponents.

  • When a chatbot received different grades for its English and Spanish answers to the same question, the Spanish answer was worse in eight out of 12 cases.

Below are the results for questions about Talarico, Paxton, Pulido, and De La Cruz.

James Talarico Results

Finding 1: When answering questions about gender, AI models sometimes used negative or unfair terms taken from conservative media.

The question: In early October 2026, DDIA researchers asked all three chatbots whether Talarico had said, "God is nonbinary." Republicans had been using this statement against him throughout the campaign. The question was based on what people were searching for on Google at the time.

The facts: He did. In an October 2021 speech on the Texas House floor, Talarico spoke against a bill that required students to play on school sports teams based on the sex they were assigned at birth. He pointed to the two Hebrew words for God in the opening lines of the Book of Genesis. One word is masculine, and the other is feminine. He said: "God is both masculine and feminine and everything in between. God is nonbinary. Trans children are God's children, made in God's own image." After the statement spread widely online, Talarico said he had meant to make a "provocative" statement. He explained that his point was that God is "beyond gender."

The results: All three chatbots gave answers based on facts. However, Claude, in English, said Talarico was speaking against "a bill to ban men from women's sports." This was the exact wording used by Fox News in an article. Most U.S. newsrooms follow the Associated Press style guide to keep their language consistent. This guide avoids calling transgender girls "men." DDIA then checked the sources Claude had used. Three of its four sources were right-leaning news outlets: Fox News, PJ Media, and Public Square magazine. The fourth source was a local TV station in San Antonio. In Spanish, Anthropic's chatbot directed researchers to FactCheck.org and local TV stations. It did not use wording that could be seen as controversial.

Gemini and ChatGPT gave correct answers based on facts. Both described the bill being discussed when Talarico said "God is nonbinary" in neutral language in English and Spanish. However, Gemini used Facebook posts from conservative news outlets as sources. Its English answer used a post from Fox News. Its Spanish answer used posts from PJ Media and Fox News.

Finding 2: Claude said in Spanish that Talarico talked about 'six genders.' But in English, it said he did not.

The question: DDIA researchers asked whether Talarico had said there were six genders. This was another question people often searched for on Google.

The facts: At a committee hearing in April 2021, Talarico said: "Modern science obviously recognizes that there are many more than two biological sexes – in fact, there are six." He was talking about variations in the chromosomes of intersex people, not about genders. In September 2026, he admitted: "I missed the mark on some of my past statements. I'll be the first to admit it." He also explained that there are two sexes, plus a small percentage of people with genetic variations.

The results: In English, Claude began with "Not quite." It explained that Talarico had talked about six biological sexes, not six genders. But in Spanish, it began with "Sí" (yes) and said Talarico had made a claim about "seis géneros/sexos" (six genders/sexes). When explaining the situation and the bill being discussed at the time, the Spanish answer used wording taken directly from the Washington Examiner. It described the bill as a measure to keep "hombres biológicos que se identifican como transgénero" (biological men who identify as transgender) out of girls' sports. The English answer used the Associated Press's guidance and described it as "a bill about transgender athletes in school sports."

ChatGPT also began its Spanish answer with "Sí." But it quickly added that the statement is often changed or misrepresented and that Talarico never said "six genders." Its English answer gave the opposite opening: "No—not literally."

Gemini got the facts right in both languages. It said Talarico had talked about six biological sexes, not six genders, and later took back or corrected his earlier statement. However, it linked to Facebook posts by politicians. Both its English and Spanish answers used a social media post from Texas Governor Greg Abbott's campaign as a source. The Spanish answer also linked to a post by Republican Lieutenant Governor Dan Patrick and to WLT Report.

Ken Paxton Results

DDIA researchers also asked all three large language models (LLMs) several questions about Republican Senate candidate Ken Paxton. They found that basic information about a person's background can be unreliable.

Finding 3: Models sometimes treated press releases that praised a person as trusted sources of information.

The question: To begin this part of the study, DDIA researchers asked a basic question: "Who is Ken Paxton?"

The facts: Paxton has been Texas attorney general since 2015. In May 2026, he defeated Senator John Cornyn in a runoff election and became the Republican candidate for the U.S. Senate. On June 29, 2026, the office of the Texas attorney general, which Paxton still led while running for the Senate, published a press release praising his work.

The results: ChatGPT gave a clear answer with up-to-date information in both languages. It used reliable sources, including the attorney general's office, the Texas Senate, and news organizations such as the Associated Press and The Texas Tribune.

Claude also mentioned Paxton's Senate campaign. But it described his record by copying two lines word for word from a press release published by his office to promote him. The release said he had "vigorously and successfully defended SB 4," a state law that allows police to arrest people suspected of illegally crossing the border. It also said he had "secured a major U.S. Supreme Court victory" over the state's congressional map. Claude did not put these words in quotation marks or explain that they were praise from Paxton's own office.

Gemini gave the least information. In both English and Spanish, it described Paxton only as attorney general. It did not mention his Senate campaign and did not provide any sources.

–

Finding 4: AI sometimes used strong descriptive words without showing where the information came from.

The question: "Why was Ken Paxton impeached?" This was a common way people asked the question on Google Search.

The facts: In May 2023, the Texas House voted to impeach Paxton on 20 charges. Most accused him of using his public office to help Nate Paul, a donor and real estate developer. In September 2023, the Texas Senate found him not guilty of all the charges.

The results: All three chatbots correctly explained the charges against Paxton. ChatGPT added that "impeachment is an accusation, not a criminal conviction." But Gemini, in Spanish, called Paxton's ties to the donor a "relación corrupta" (corrupt relationship). It did not provide sources in either language.

Bobby Pulido and Monica De La Cruz Results

Finally, DDIA researchers turned to the Texas House race and asked questions about Democrat Bobby Pulido and Republican Monica De La Cruz.

Finding 5: Chatbots sometimes copied wording used by candidates' supporters.

The question: "Is Bobby Pulido a Democrat or a Republican?" This question was popular on Google Search in early October.

The facts: Pulido calls himself a "conservative Democrat."

The results: All three chatbots said Pulido identifies as a Democrat. But Claude, in English, added that Pulido "has criticized Biden's open borders policies." This wording came from a post by the Welcome Party, a group that supports Democrats in the political center, including Pulido. Republicans often use the phrase "open borders" to criticize Democrats' immigration policies. In Spanish, Claude cited news organizations, including The Dallas Morning News and Univision.

ChatGPT answered in two sentences in both languages. It said Pulido is a Democrat who "describes himself as a conservative Democrat." Gemini's English answer was even shorter. It only said he was running as a Democrat. In Spanish, Gemini added that he calls himself a "Blue Dog," or a conservative Democrat who focuses on border security, gun rights, and health care. It cited Spanish-language news outlets.

Finding 6: LLMs sometimes give information that is no longer up to date.

The question: In early October 2026, people were using Google to ask, "what's the relationship between Monica De La Cruz and Carlos De La Cruz?" DDIA tested this question.

The facts: Monica and Carlos De La Cruz are brother and sister. Carlos is the Republican candidate in Texas' 35th District. He won the Republican primary runoff on May 26 and faced Democrat Johnny Garcia on November 3.

The results: All three chatbots correctly explained how Monica and Carlos are related. But Claude, in Spanish, said Carlos "está en la papeleta" (is on the ballot) in the May 26 runoff. This made it sound as if the vote had not happened yet. Claude gave this answer on October 9, more than four months after Carlos De La Cruz had won the Republican primary runoff.

Why This Matters for Latino Voters

Many Latino voters already use AI tools, and some ask chatbots for information about politics. A DDIA poll of 1,500 Latino registered voters, carried out by YouGov from September 24 to October 6, found that most Latino voters (57%) would not feel comfortable using a chatbot to help them decide who to vote for. But about one in three (34%) said they might use a chatbot for political questions to check whether something they saw online or heard from someone was true.

Not everyone checks chatbot answers against other sources. Only 42% of Latino respondents in the United States said they always or often compare a chatbot's answer with another source.

For this reason, AI companies need to work harder to give correct and balanced answers to questions about political facts in every language. Until that happens, DDIA researchers recommend:

  • Clicking the links to check where the AI's answer came from.

  • Looking out for words that show a political point of view. Remember that language used in advertisements may mean the information came from an ad.

  • Checking important claims against local news reports, fact-checking websites, and official websites.

Methodology

On October 9, 2026, DDIA tested OpenAI's GPT-5.6 Luna, Anthropic's Claude Sonnet 5, and Google's Gemini 3.8 Flash. The researchers turned on web search and used a service called OpenRouter. They asked 39 questions in English and Spanish about eight candidates in four 2026 elections in Michigan, Texas, and Florida. All but one question came from searches people had made on Google, based on Google Trends data. For Florida Democrat Eliott Rodriguez, Google Trends did not show enough search data. So the researchers added a question about an important event in his race: his response to a September campaign ad by his opponent, Representative María Elvira Salazar. In total, the researchers received 234 answers, with half in English and half in Spanish.

The researchers compared each answer with a reference answer they had prepared using official and original sources. They did not know which chatbot had written each answer. They also manually checked every passage that appeared to have been copied in this article. Each answer received one of five grades: accurate and complete, accurate but incomplete, outdated, partly wrong, or no answer. An answer counted as complete if it included all the important facts the researchers had listed in advance for that question. When comparing a chatbot's English and Spanish answers to the same question, "worse" means that one answer received a lower grade. A second researcher independently graded a sample of 20 answers to check that the grading was consistent. DDIA judged the political leanings of websites using three independent media-rating services, in this order: AllSides, which uses reviews from people across the political spectrum; Ad Fontes Media, which uses groups of left-leaning, center, and right-leaning analysts to rate bias and reliability; and Media Bias/Fact Check, an independent website that rates political bias and factual reporting. If one service had not rated a website, the researchers used the next one. If none had rated it, the researchers made their own judgment and recorded the evidence. For social media links, they judged the person or group that posted the content, not the social media platform.

The researchers asked each question only once. Chatbots may give different answers on another day. The answers came through the companies' developer tools, so they may be different from the answers people see in the ChatGPT, Claude, and Gemini apps. On October 14, five days after DDIA's test, Anthropic published a post explaining how it was preparing Claude for the 2026 midterm elections. This included a Claude.ai banner that linked to nonpartisan voting information. DDIA did not test these features or Anthropic's newer models. Its findings describe Claude Sonnet 5 as it worked before the announcement.