Four Chatbots. 3,100 Prompts. 90% Failure on Election Questions Six Months Before a Midterm.
- 90%Failure rate of ChatGPT / Gemini / Claude / Grok on election questions across 3,100+ Forum AI prompts (Bloomberg, May 20, 2026).
- 45%Of AI-assistant news answers contained significant errors in the October 2025 EBU/BBC global study across 22 outlets and 18 countries.
- 28%+Baseline false-claim repetition rate on controversial news in NewsGuard's January 2026 quarterly AI False Claims Monitor.
- 15%Anthropic Claude's April 2026 rate of citing Russian state-affiliated media — up from 4% baseline (NewsGuard).
- $852BOpenAI's late-2025 private-market valuation — no binding accuracy obligation, no election-misinformation liability, no federal audit.
- 0Federal election-misinformation regulations specific to consumer chatbots, six months before the November 3, 2026 general election.
On May 20, 2026 — roughly six months before the November 3 midterm election — Bloomberg Technology published the results of a 3,100-prompt audit of the four chatbots most Americans actually use. ChatGPT, Google Gemini, Anthropic Claude, and xAI's Grok were asked election questions across the spectrum: polling-place hours, voter-registration rules, candidate stances, ballot deadlines, ID requirements. The Forum AI study's headline finding: the four chatbots failed on accuracy, bias, or source selection roughly 90 percent of the time.
This is not an outlier finding. In October 2025, the European Broadcasting Union and the BBC ran the largest-ever international study of AI news accuracy — 22 public-service broadcasters, 14 languages, 18 countries, roughly 3,000 evaluated responses. The result: AI assistants misrepresent news content 45 percent of the time, regardless of language or territory. NewsGuard's January 2026 quarterly tracker put the baseline false-claim repetition rate on controversial news at 28 percent and rising. In April 2026, NewsGuard documented that Anthropic's Claude was citing Russian state-affiliated media 15 percent of the time — up from a 4 percent baseline — and citing Iranian state media for the first time.
Hundreds of millions of voters now use these tools as their primary information layer. Yet the FTC under Chair Andrew Ferguson (R) has explicitly said it will not be a “general AI regulator.” The FCC under Chair Brendan Carr (R) is focused on broadcast licensing. Senate Commerce under Sen. Ted Cruz (R-TX) has bandwidth for kids-safety bills but not for election-misinformation oversight. And the four companies behind the products — OpenAI (~$852B valuation), Anthropic (~$900B), Alphabet, and xAI — have no binding accuracy obligation, no audit requirement, and no liability for election misinformation generated at scale.
The Bloomberg-reported Forum AI audit graded four chatbots across three dimensions: factual accuracy, evident partisan bias, and quality of cited sources. A response counted as “failed” if it tripped any of the three. Across 3,100+ prompts, all four chatbots tripped one or more in roughly 90 percent of responses.
Specific failure modes documented across the Bloomberg piece and corroborating outlets: ChatGPT and Gemini both confidently answered with election dates that no longer applied. Claude cited Russian state-affiliated media in 15 percent of foreign-policy answers. Grok, per a May 13, 2026 SF Standard piece, was the only chatbot of the four that would directly recommend candidate votes for the California June 2 primary — while ChatGPT, Claude, and Gemini declined to recommend. The Brennan Center has published a separate analysis documenting Grok's history of spreading false election information including incorrect ballot deadlines and disqualification claims.
The Bloomberg / Forum AI audit lands on top of a year of cross-corroborated research. On October 21, 2025, the European Broadcasting Union and the BBC published the largest international study of AI news accuracy ever conducted. The findings:
- 45 percent of AI-assistant news responses contained at least one significant error.
- 81 percent contained at least some error.
- Google Gemini performed worst — 76 percent significant-issue rate.
- The error pattern was consistent across 14 languages and 18 countries — not a US-specific or English-language phenomenon.
- The four chatbots tested: ChatGPT, Microsoft Copilot, Google Gemini, and Perplexity.
“AI's systemic distortion of news is consistent across languages and territories.”
European Broadcasting Union joint study with the BBC, October 21, 2025
NewsGuard's ongoing AI False Claims Monitor caught a specific, dangerous pattern in April 2026. Anthropic's Claude — the chatbot the company markets as “safer” and “more honest” than its competitors — was citing Russian state-affiliated media (RT, Sputnik, TASS) in 15 percent of its foreign-policy answers. The baseline a year earlier had been 4 percent.
NewsGuard separately documented Claude's first instance of citing Iranian state media — almost certainly tied to a 2026 Iranian state-media campaign that produced hundreds of English-language posts mimicking Western news outlets, posts that then got indexed and recalled by Claude.
Hundreds of millions of Americans now ask chatbots for election information. The chatbots themselves disclaim accuracy in fine print but answer confidently in the main response.
If a TV broadcaster pushed election misinformation in 90 percent of segments six months before a vote, the FCC would act within hours. When four chatbots used by hundreds of millions do the same, the federal response so far is a kids-safety bill from the Cruz/Schatz/Curtis/Schiff working group and two Gillibrand letters to FTC Chair Ferguson asking the agency to step in. The agencies, by their own admission, are not stepping in.
The legislative response is, so far, narrow. Sen. Ted Cruz (R-TX), chair of Senate Commerce, introduced the CHATBOT Act in May 2026 along with Sens. Brian Schatz (D-HI), John Curtis (R-UT), and Adam Schiff (D-CA). The bill regulates AI-chatbot use by minors — disclosure, parental controls, transparency on AI companion bots. Sen. Josh Hawley (R-MO) has a companion ban on AI companion bots for minors. None of the pending bills specifically regulate election-information accuracy.
Sen. Kirsten Gillibrand (D-NY)has written two letters to FTC Chair Andrew Ferguson (R) — February 9, 2026 and March 12, 2026 — pressing the agency to enforce against deceptive AI chatbot claims. Ferguson's public position: existing FTC authority covers deceptive practices and the agency will use it case by case, but it will not become a “general AI regulator.”
FCC Chair Brendan Carr (R) has called Silicon Valley AI bias “deeply concerning” and proposed February 2026 action against “misinformation” broadcasters, but FCC authority does not extend to chatbots — the same Communications Act provisions that govern broadcast stations do not cover internet-delivered AI services. Carr's rhetoric and the FCC's actual legal reach do not match up.
Introduced the CHATBOT Act with Sens. Schatz (D-HI), Curtis (R-UT), and Schiff (D-CA): disclosure, parental controls, transparency for AI companion bots aimed at minors.
Source: Senate Commerce committee press release with bill text and bipartisan cosponsor list.
Two letters to FTC Chair Andrew Ferguson (R) pressing the agency to enforce against deceptive AI-chatbot claims — formal congressional pressure on the federal-regulator response gap.
Source: Senator Gillibrand's official PDF of the letter text.
OpenAI's position, articulated in Fox News Media reporting: GPT-5 reduces measured political bias by roughly 30 percent vs. GPT-4 per internal evaluation. Anthropic's position: Claude is safety-trained to refuse politically loaded recommendations, and the Russian-citation pattern is a search-index artifact being remediated. Google: Gemini's issues are being addressed in successive model releases; the 76% significant-issue rate reflects a specific dataset and answer format.
xAI's position, more visibly marketed: Grok is the “only non-woke” chatbot, designed explicitly to answer politically loaded questions where competitors refuse. The Brennan Center documents that this marketing positioning corresponds to a measurably higher rate of election-specific misinformation — including specific false claims about ballot deadlines, polling-place hours, and disqualification rules.
Big Tech and AI are working overtime to influence elections. We see Google's Gemini, ChatGPT, all of them, pushing biased answers about American politics. The American people deserve honest information — not Silicon Valley propaganda.
Paraphrased commentary · not a verbatim post
Composite paraphrase of Trump's repeatedly stated position on AI election bias, cross-referenced via Fox News Media coverage of his February 2024 Gemini criticism and subsequent posts.
OpenAI is essentially captured by Microsoft. ChatGPT is biased. The American people will see through it.
Paraphrased commentary · not a verbatim post
Paraphrase of the Trump posture on AI captured by tech giants, cross-referenced via NPR coverage of his Truth Social posts on Big Tech bias.
The political clock is running. November 3, 2026 is fewer than 175 days from the publication of this story. The Forum AI study that produced the 90 percent finding will likely be re-run in the fall. The CHATBOT Act may pass. The Gillibrand letters may move the FTC. NewsGuard's tracker will keep publishing. None of those address the underlying mechanism: four chatbots answering hundreds of millions of election questions with no obligation to be right, no audit they have to pass, and no liability for the consequences when they are wrong.
“The political ideology that permeates Silicon Valley is finding its way into algorithms.”
FCC Chair Brendan Carr (R), on AI bias in chatbot outputs
90 percent failure rate on election questions across ChatGPT, Gemini, Claude, and Grok (Forum AI / Bloomberg, May 20, 2026). 45 percent material-error rate across all news topics (EBU/BBC, October 2025). 28 percent+ false-claim repetition baseline (NewsGuard, January 2026). 15 percent of Claude responses cite Russian state media (NewsGuard, April 2026).
The FTC under Ferguson will not be the regulator. The FCC under Carr has no jurisdiction. Senate Commerce under Cruz is on kids-safety. The four companies — OpenAI, Anthropic, Alphabet, xAI — have no binding accuracy obligation 175 days before the midterm.
When a TV station misinforms voters, the FCC acts within hours. When four chatbots used by hundreds of millions misinform them at scale, the federal government has, so far, sent two letters.