Leading Up to the U.S. Midterm Elections, DDIA Asked ChatGPT, Claude, and Gemini about Key Races in Texas, Florida, and Michigan. We Found Partisan Language, Outdated Facts, and Discrepancies Between English and Spanish
Texas, Florida, and Michigan findings are available in greater detail here.
By: Rafael Olavarría, Research Consultant, DDIA
Latino voters in the United States are increasingly turning to AI to stay informed about politics and democracy. As of October 2026, 22% of Latino voters reported using ChatGPT for media consumption on a daily basis, and 45% reported broadly using AI on a regular basis. Though most are not yet using AI to make decisions about which candidates to vote for, it is clear chatbots are, in many ways, the new search engines of our time.
To assess how three leading AI chatbots – OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini – perform when answering questions voters were frequently searching for online, on October 9, 2026, DDIA asked each model 39 questions about eight candidates in four U.S. midterm races: the U.S. Senate races in Michigan and Texas, and the House races in Texas' 15th and Florida's 27th districts.
Most of the questions the team posed were based on what people were searching for on Google. Each question was asked in English and in Spanish, and generated a total of 234 answers. Researchers graded each answer on a five-point scale: accurate and complete, accurate but incomplete, outdated, partly wrong, or didn’t answer.
Though most answers (77%) were accurate and complete, DDIA observed a few patterns worth noting and addressing, including:
Partisan language presented as fact: In many cases, chatbots, mostly Claude, frequently copied politically charged language, sometimes word for word, from partisan sources. They repeated expressions used to praise candidates or attack opponents without attribution or context. Sources the models cited included partisan news outlets, campaign websites, and candidates’ official biographies.
Outdated information presented as current: Chatbots sometimes presented outdated information as if it was current. This problem occurred twice as often in Spanish as in English.
Wrong, incomplete and made-up answers: 37 of 234 answers were partly wrong or left out essential information, 22 of them in Spanish. The three answers DDIA identified as having made-up information were all in Spanish.
Unreliable sources treated as credible reporting: Chatbots relied on social media posts, partisan outlets, and opinion columns to provide answers without consistently distinguishing them from independent news reporting or official data.
Spanish-language answers relying heavily on English-language journalism: Rather than drawing from Spanish-language news coverage, chatbots largely translated information published by English-language U.S. outlets. They rarely cited major Spanish-language news organizations such as Univision and Telemundo.
Claude showing the strongest partisan language and the widest gap between English and Spanish: Of the three chatbots tested, Claude stood out for its frequent use of politically loaded language and its reliance on partisan outlets. When its English and Spanish answers differed in quality, the Spanish response was worse 91% of the time.
Note: On Oct. 14, five days after DDIA’s test, Anthropic published a post about how the company was preparing for the 2026 midterms. It said people asking Claude about politics “should get accurate and evenhanded answers that don’t favor any particular political viewpoint.” The company also reported that its newer 5.5 models scored 99% on the even-handedness evaluation. It also described a Claude.ai banner linking to nonpartisan voting resources. DDIA did not test these features or the newer models; its results reflect Claude Sonnet 5 before the announcement.
OpenAI and Google published their election commitments before DDIA’s test. In a May 27, 2026 post, OpenAI said it monitors its models “to keep ChatGPT’s responses politically neutral” and aims to keep ChatGPT “objective by default.” Google’s Sept. 9, 2026, midterms post focused on connecting users, including in the Gemini app, with official voting information and real-time results from The Associated Press. It did not address whether Gemini’s answers should be politically neutral. None of the three companies’ posts mentions Spanish.
In other publications, however, all three companies acknowledge that their models can perform differently across languages. OpenAI said in its October 2025 political bias evaluation that bias “can vary across languages and cultures,” and Anthropic’s July 2026 research found Claude “may also be more closely matching our intended behavior for some languages than others.” Google’s September 2026 model card for Gemini 3.8 Flash, the version DDIA tested, noted that its “safety performance across non-English languages regressed slightly” from the previous version.
Finding 1: Loaded or partisan language presented as fact
DDIA checked all 234 answers for wording that echoed one side’s talking points. In 24 of them, chatbots adopted politically loaded language as their own, without quotation marks or attribution. Claude accounted for 18 of those answers, compared with five from Gemini and just one from ChatGPT. In 14 of Claude’s 18 cases, researchers found the wording copied, almost word for word, from a page Claude linked to in that same answer.
In one case, Claude closely reproduced loaded language from a Fox News article describing Michigan Democratic Senate candidate Abdul El-Sayed’s ties to Jeffrey Soros, nephew of billionaire George Soros.
In another, Claude described now former Texas Republican Senate candidate Ken Paxton using promotional language from a press release issued by his own office. It also praised Florida Republican Rep. María Elvira Salazar, who was seeking reelection, using language from her official biography, saying she is "committed to acting in defense of individual rights and liberties."
This behavior extended beyond candidate biographies to policy debates. Claude claimed that then Texas Democratic House candidate Bobby Pulido "has criticized Biden's open borders policies," repeating language from the Welcome Party, a group supporting centrist Democrats. The phrase "open borders" is commonly used by Republicans to attack Democrats on immigration.
Gemini also adopted partisan language, sometimes introducing it only in Spanish. It linked El-Sayed’s political positions to "la izquierda radical" (the radical left) and described Paxton’s ties to a donor as a "relación corrupta" (corrupt relationship). Neither label appeared in Gemini’s English-language answers.
Republicans seemingly benefited more often from this pattern, though the imbalance partly reflects the questions tested: many of the most-searched queries extracted from Google Trends were based on or were related to attacks against Democratic candidates. Of the 24 answers with loaded language, 14 used wording associated with Republican talking points, such as calling transgender girls “men” or Biden’s policies “open borders,” or repeating Republican candidates’ self-praise. Seven used wording associated with Democrats, such as describing the Miami area’s “affordability crisis” or Mike Rogers’ AIPAC backers as a “donor class.”
Loaded language presented as fact was also more common when chatbots addressed politically contentious subjects. Loaded answers showed up even when the questions did not contain loaded words (Claude used “open borders” when asked simply whether Pulido was a Democrat or a Republican). Thirteen percent of answers about attacks, controversies, and claims requiring verification adopted partisan wording, compared with 6% of answers to basic factual questions, such as who a candidate is or which office they are seeking.
Finding 2: Outdated News Presented as Current
Outdated information presented as current information was more common in Spanish-language answers. Of the 234 answers from all three chatbots, 16 included outdated information presented as current, and 11 of those were in Spanish. Claude produced the most outdated responses, the majority of them in Spanish.
Claude told Spanish-speaking users, for example, that Paxton was still competing in his party’s Senate primary, more than four months after winning the May 26 runoff. It also claimed that Carlos De La Cruz, brother of Rep. Monica De La Cruz, a Republican who was seeking reelection in Texas, was still "on the ballot" in the same runoff, even though he had already won it.
In Florida, Claude told Spanish-speaking users that Democrat Eliott Rodriguez had "announced on Tuesday" that he was running for Congress, repeating language from news articles published in March 2026, when he launched his campaign.
Gemini, meanwhile, identified Paxton only as Texas’ attorney general in both languages, failing to mention his U.S. Senate campaign.
Finding 3: Made-up Facts and Weaker Answers in Spanish
The three answers DDIA identified as including made-up facts throughout the experiment were in Spanish, two from Gemini and one from Claude.
Gemini, for example, falsely identified Salazar’s first husband as "Arturo Salabarria," a name researchers found in no record connected to Salazar. The congresswoman has publicly identified her first husband as Arturo Cruz.
Gemini also said Bobby Pulido appeared in the telenovela "Las tontas no van al cielo." He never did. His telenovela appearances include "Fuego en la sangre" and "Qué pobres tan ricos." It also gave his birth year as 1971; he was born in 1973.
Claude, meanwhile, described Michigan Republican Mike Rogers as "senador electo" (senator-elect), though he lost his 2024 Senate race and was running again in 2026. Mike Rogers has since lost the election to Abdul El-Sayed.
Separately, 16 answers included at least one false statement, 10 of them in Spanish, and 21 answers left out essential information, 12 of them in Spanish.
DDIA compared each chatbot’s English and Spanish answers to the same question, 117 pairs in all (39 questions x 3 chatbots). In 24 of those pairs, the two answers got different grades. In 18, the English answer was better; in six, the Spanish one was.
Answers offered about Florida showed the widest discrepancies between languages. In nine of the 10 pairs where answer quality differed, the Spanish response was worse.
One example involved Florida Democratic congressional candidate Eliott Rodriguez. When asked "Who is Elliot Rodriguez?"—using the misspelling most common in Google searches—Claude correctly identified the candidate in English. In Spanish, however, it claimed it could not determine whom the researchers meant, despite correctly identifying Rodriguez in other Spanish-language responses.
Finding 4: Questionable or Biased Sources
Partisan and questionable websites: Claude cited right-leaning sources more often than the other two models, with 47 links, including WLT Report, which Media Bias/Fact Check, an independent website that evaluates news outlets, classifies as questionable and low in factual reporting. Gemini also cited WLT Report and Conservative Underground News, another outlet rated as questionable. Claude was more than twice as likely to adopt partisan language when citing partisan sources: 44% of those answers (seven of 16) used politically loaded wording, compared with 18% of its other answers (11 of 62). Gemini and ChatGPT cited partisan outlets in nine answers combined, but none adopted their loaded language. DDIA traced several of Claude's expressions directly to right-leaning sources it cited, including Fox News and the Washington Examiner. In another answer, about El-Sayed's mother, Claude copied passages from The Western Journal, a conservative outlet that says it "upholds traditional Christian values as articulated in the Bible." The chatbots cited more sites rated as left-leaning (256 links, mostly mainstream outlets such as CBS News and The Texas Tribune) than right-leaning (79). But 30 of the 33 links to openly partisan or opinion outlets went to right-wing sites, and the loaded wording DDIA traced came from right-leaning pages.
Opinion and political self-promotion presented as reporting: Claude blurred the distinction between independent journalism, opinion, and promotional material. It introduced a Detroit News opinion column with the phrase "As the Detroit News points out," treating a columnist's opinion as the newspaper's reporting. It also repeated language from Salazar's official biography, including outdated information about her husband, using two websites that had republished it.
Social media posts used as sources: Gemini cited social media posts 54 times, primarily YouTube videos and Facebook posts, accounting for nearly all social media citations in the study. Most of these answers were still accurate, and linking to a candidate’s own official page can be a reasonable way to answer a question about them. But some of these links went to posts by the candidates’ political opponents: Gov. Greg Abbott’s campaign and Lt. Gov. Dan Patrick on Democrat James Talarico, and Rep. Nancy Mace on Democrat Abdul El-Sayed. This meant voters attempting to verify an answer could land on a political opponent’s post instead of a news report. Gemini also provided no sources at all in 10 answers, making independent verification more difficult.
Finding 5: In Spanish, the Chatbots Mostly Translated English-Language News
All three chatbots rarely cited Spanish-language journalism when answering questions in Spanish. Of the 404 links cited in Spanish-language responses, only 48 (12%) led to Spanish-language news outlets. Just eight pointed to Univision or Telemundo, the two largest Spanish-language television networks in the United States.
Instead, chatbots relied heavily on English-language U.S. news sources, which accounted for 164 links (41%) out of 404. Spanish-language outlets based in the United States received just 18 citations, fewer than the 23 links to Spanish-language outlets abroad, primarily in Mexico and Argentina.
Claude cited Fox News more often in Spanish (six times) than Univision and Telemundo combined (three times). The preference for English-language sources extended beyond journalism: of the 36 Wikipedia links included in Spanish-language answers, 34 directed users to English-language pages.
How Each Model Performed
ChatGPT had the fewest problems. Researchers flagged just one answer for partisan language and two for outdated information. ChatGPT produced no fabricated facts and cited no social media posts. Its responses tended to be shorter and included fewer sources.
Claude stood out for partisan language, copied wording, and outdated information. Claude produced lengthy, heavily sourced answers but was the most likely to reproduce politically loaded language, often nearly word for word, from pages it linked to, without quotation marks or attribution. It also had the most answers containing outdated information. When its English and Spanish responses differed in quality, the Spanish answer was worse 91% of the time.
Gemini relied heavily on social media and sometimes provided no sources at all. Although most of its answers were accurate, it cited social media posts 54 times, accounting for 17% of all its links, compared with just two social media citations from Claude and none from ChatGPT. Gemini provided no sources in 13% of its answers, more than the other two chatbots combined. It also generated two of the study’s three answers containing fabricated facts, both in Spanish.
Questions For Consideration
OpenAI, Anthropic and Google have all published commitments on how their chatbots should handle elections, partisan topics and accuracy. DDIA’s test shows how well those standards held up in practice, in English and in Spanish. The questions below can serve as a bridge between what the companies say their chatbots should do and what DDIA found they did:
1. Why did Claude adopt partisan wording despite its own standard against partisan language, and what helped ChatGPT avoid it?
OpenAI’s December 2025 Model Spec says its models should, by default, assume an objective point of view and avoid “subjective terms unless quoting directly or citing specific sources.” Anthropic says in its November 2025 political standards and its January 2026 constitution that Claude should “adopt neutral terminology over politically-loaded terminology where possible.” Google’s AI Principles commit to safeguards that “avoid unfair bias,” but neither they nor its September 2026 midterms post address whether Gemini’s answers should avoid favoring a party or candidate.
In DDIA’s test, ChatGPT used loaded wording in one answer and Gemini in five, all in Spanish. Claude did so in 18.
2. Are U.S. election safeguards tested and monitored in Spanish?
All three companies acknowledge that their models can perform differently across languages. OpenAI said in its October 2025 bias evaluation that bias “can vary across languages and cultures,” and that the evaluation began with “U.S. English interactions.” Anthropic’s July 2026 research found Claude “may also be more closely matching our intended behavior for some languages than others,” and its June 2026 system card for Claude Sonnet 5, the version DDIA tested, reports safety tests in seven languages, not including Spanish. Google’s September 2026 model card for Gemini 3.8 Flash, the version DDIA tested, said its “safety performance across non-English languages regressed slightly,” and its February 2026 responsibility report says it phases rollouts to allow “safety considerations for different languages and regions.”
But none of their U.S. election announcements mentions Spanish, the most common language other than English spoken in U.S. homes. When Claude’s English and Spanish answers got different grades, the Spanish one was worse in 10 of 11 questions. For ChatGPT, it was worse in four of seven.
3. Why do Spanish-language responses to Spanish-language questions posed in the United States draw so little from U.S. Spanish-language outlets, even when those citations are the most relevant?
None of the companies explains how its models choose which news outlets to cite.
In Spanish answers, ChatGPT cited Univision and Telemundo four times in 68 links. Claude cited them three times in 168 links, while citing Fox News six times. Gemini cited them once in 168 links, while linking to social media posts 31 times.
4. When a chatbot searches the web, what factors does it take into account when selecting the most recent content to feature in its responses? How can the models be trained to avoid grounding responses in old or outdated information when we know more up-to-date information is readily available?
Anthropic said in an April 2026 update that Claude has a “knowledge cutoff,” but “when web search is enabled, Claude can find and relay up-to-date information.” Google says in its February 2026 report that it uses Search “to factually ground responses that require fresh or authoritative information.” OpenAI’s documents reviewed by DDIA do not address the issue.
All three searched the web in DDIA’s test. Claude still gave nine outdated answers, seven in Spanish, compared with five from Gemini and two from ChatGPT.
5. How can chatbots be trained to differentiate between opinion pieces and news pieces, even when both are published on legitimate news sites?
Anthropic’s constitution says Claude should show “appropriate skepticism toward content from low-quality or unreliable websites,” and lists guidance on how to “evaluate and weight search results with differing levels of reliability” as possible future work. Google tests its models’ factual accuracy with benchmarks it describes in its February 2026 report. But none of the documents DDIA reviewed from the three companies addresses opinion pieces published by reputable outlets.
This represents a potential blind spot for the bots. As mentioned above in our study, we found that Claude drew information from a Detroit News opinion column and introduced it with “As the Detroit News points out,” presenting a columnist’s view as the newspaper’s reporting.
Looking Ahead
Exit polls from the midterm elections showed that more voters see AI as doing more harm than good to society, and that majorities in Senate battleground states oppose building a data center in their own community. Our own DDIA poll conducted with YouGov between September and October showed that Latinos are more worried about how fast AI is developing (38%) than are excited (20%), while another 24% said they had mixed feelings. Voters see a clear need for stronger safeguards, with 78% of Latino registered voters agreeing there is a pressing need for stricter rules and oversight of AI, including 72% of Latino Republicans and 83% of Latino Democrats, and that 51% oppose building a data center in their own community.
The new U.S. Congress coming into office in January 2027 will have the legitimacy to propose a new paradigm where AI development, safety and ethical standards go hand in hand, and where communities have a seat at the table.
DDIA’s goal in building a healthier information ecosystem for Latino communities includes shedding light on what is working and what needs improvement as these systems develop. Our team is committed to testing frontier chatbots in both languages. Findings of this research and future experiments will support stronger terms of service for the platforms, and stronger oversight by the new electeds paving the way for a better future for all online.
Download PDF:
DownloadMethodology
On Oct. 9, 2026, DDIA tested OpenAI's GPT-5.6 Luna, Anthropic's Claude Sonnet 5 and Google's Gemini 3.8 Flash, with web search on, through a service called OpenRouter. Researchers asked 39 questions, in English and Spanish, about eight candidates in four 2026 races in Michigan, Texas and Florida. All but one of the questions were based on what people searched for on Google, using Google Trends. For Florida Democrat Eliott Rodriguez, Trends showed too little search data, so researchers added one question about a major news event in the race: his response to a September campaign ad by his opponent, Rep. María Elvira Salazar. In total, researchers received 234 answers, half in English and half in Spanish.
Researchers graded each answer against a reference answer written from official and primary sources, without knowing which chatbot wrote it, and manually checked every copied passage in this story. Each answer got one of five grades: accurate and complete, accurate but incomplete, outdated, partly wrong, or didn’t answer. An answer counted as complete if it included the key facts researchers listed in advance for that question.
When comparing a chatbot’s English and Spanish answers to the same question, “worse” means a lower grade. A second researcher independently graded a sample of 20 answers to check consistency.
DDIA rated websites' political leaning using three independent media-rating services, in this order: AllSides, which rates outlets' bias using reviews by people from across the political spectrum; Ad Fontes Media, which rates bias and reliability using panels of left, center and right analysts; and Media Bias/Fact Check, an independent website that rates bias and factual reporting. When one service hadn't rated a site, researchers used the next; when none had, researchers made their own judgment and recorded the evidence. Social media links were rated by who posted them, not by the platform. DDIA did not check whether individual news sites limit AI tools’ access, which could affect which outlets the chatbots cited
Each question was asked once, and chatbots can answer differently on another day. Answers came through the companies' developer tools and may differ from what users see in the ChatGPT, Claude and Gemini apps.
