Skip to content
Theorem Weekly TW markTWTheoremWeekly

The evidence behind the week in machine learning.

Model updates

Forum AI study finds mainstream chatbots misstate election facts nearly 40 percent of the time

Research by Forum AI and NewsGuard reveals high rates of factual errors and partisan bias in major AI models during midterms.

Key points

  • Forum AI found almost 40 percent of election answers from mainstream chatbots contained at least one factual error.
  • NewsGuard reported that mainstream chatbots cite partisan "pink slime" websites nearly 50 percent of the time for midterm queries.
  • Google's Gemini incorrectly stated Pennsylvania's mail-in ballot deadline was November 3 instead of October 27.
Forum AI study finds mainstream chatbots misstate election facts nearly 40 percent of the time
Depth
40 percentForum AI found almost 40 percent of election answers from mainstream chatbots contained at least one factual error · reported
50 percentNewsGuard reported that mainstream chatbots cite partisan "pink slime" websites nearly 50 percent of the time for midterm queries · reported
1 outlet · newest source 308h ago

Forum AI research indicates that nearly 40 percent of responses from mainstream artificial intelligence chatbots regarding the US midterm elections contained factual errors.

Evidence: reportedHere’s Why You Shouldn’t Use AI to Vote

aifleet-wired

What was released or claimed

Forum AI researchers tested major models including ChatGPT, Claude and Gemini on election-related questions. The study found significant inaccuracies, such as Claude confusing a North Carolina congressman with a former school superintendent of the same name. Google's Gemini also provided incorrect information about ballot deadlines in Pennsylvania.

Who it affects

US adults are the primary group affected, with Pew Research reporting that 8 percent have used AI tools to find information on candidates and policies ahead of the midterms. Approximately 50 percent of surveyed US adults have used an AI chatbot at some point.

What is still unknown

It remains unclear how widespread the use of AI for voting decisions is beyond the surveyed 8 percent. The extent to which these errors influence actual voting behavior or candidate preference among the general public is not measured by the cited research.

Evidence: reportedHere’s Why You Shouldn’t Use AI to Vote

aifleet-wired

So what

NewsGuard research highlights that chatbots rely heavily on problematic online information sources. Mainstream chatbots cite "pink slime" websites, which appear as mainstream news but are funded by political or ideological groups, nearly 50 percent of the time when asked about candidates and policies.

Evidence: reportedHere’s Why You Shouldn’t Use AI to Vote

aifleet-wired

What the paper shows

The Forum AI study demonstrates that current large language models struggle with timely, specific factual retrieval for political events. The 40 percent error rate suggests that reliance on training data and web search integration is insufficient for high-stakes civic information. NewsGuard's findings further indicate that models are not effectively filtering out ideologically motivated content, with ChatGPT citing such sites in over 70 percent of searches while Grok performed better at 30 percent. This reveals a gap between model capabilities and the accuracy required for reliable civic assistance.

Evidence: reportedHere’s Why You Shouldn’t Use AI to Vote

aifleet-wired

Why it matters

  • Researchers can observe how "pink slime" sites infiltrate model outputs, indicating failures in retrieval-augmented generation filtering.
  • The high error rate in factual queries suggests current evaluation benchmarks for political safety may be insufficient.
  • Discrepancies between models show that vendor-specific tuning significantly impacts reliability on time-sensitive civic information.
  • The reliance on partisan sources highlights the difficulty of maintaining neutrality in open-web grounded responses.

Topics: AI safety, election integrity, large language models, misinformation, Fact-checking

Updated

Then / now / next

Then

AI chatbots were generally viewed as neutral information sources for general queries.

Now

Research shows they frequently cite partisan websites and contain factual errors on election topics.

Next

Researchers will monitor how model updates address these retrieval and filtering limitations in future elections.

Story timeline

  1. Then AI chatbots were generally viewed as neutral information sources for general queries.
  2. 03 Nov 2026 Here’s Why You Shouldn’t Use AI to Vote
  3. Now Research shows they frequently cite partisan websites and contain factual errors on election topics.
  4. Next Researchers will monitor how model updates address these retrieval and filtering limitations in future elections.

The strongest doubt

Counterpoint from the sources

OpenAI directed inquiries to a blog outlining its approach to elections in 2026, rather than addressing the immediate midterm findings.

For readers

Researchers

  • Study shows 40 percent error rate in mainstream models on election queries.
  • NewsGuard data indicates 50 percent citation rate of partisan sources.

Policy

  • AI tools may misinform voters on critical deadlines and candidate details.
  • Disinformation risks increase as models cite ideologically funded websites.

Consumers

  • Verify AI-provided election information with official sources before acting.
  • Be aware that chatbots may confuse candidates or provide outdated deadlines.

Numbers in context

40 percentreported50 percentreportedCircle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.

Ask the article

Local keyword search of this article; no generated answers.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Questions readers ask

Which chatbot cited the fewest partisan websites according to NewsGuard?

xAI's Grok cited "pink slime" sites in 30 percent of cases, performing better than other tested models.

What specific error did Google's Gemini make regarding Pennsylvania?

Gemini stated the deadline to request a mail-in ballot was November 3, while the correct deadline was October 27.

How many US adults have used AI tools for midterm information?

Pew Research reported that 8 percent of surveyed US adults used AI tools to find information on candidates and policies.

Connected stories

This storyOpenAI and Anthropic Deploy Election Misinformation TeamsOpenAI and Anthropic Deploy ElecCan ERNIE 5.5 Flash Base match DeepSeek V4 on open benchmarks?Can ERNIE 5.5 Flash Base match DDoes AI-Generated Code Create Fragile SaaS Systems?Does AI-Generated Code Create FrCan LLMs Decode Unpublished Logic Notebooks Faster Than Human Scholars?Can LLMs Decode Unpublished Logi

Reading receipt

  • Forum AI found almost 40 percent of election answers from mainstream chatbots contained at least one factual error.
  • NewsGuard reported that mainstream chatbots cite partisan "pink slime" websites nearly 50 percent of the time for midterm queries.
  • Google's Gemini incorrectly stated Pennsylvania's mail-in ballot deadline was November 3 instead of October 27.

Still open: How do these error rates compare to human information reliability?

Still open: What specific mitigation strategies are vendors implementing between now and 2026?

Time spent: 0:00 · Depth: Full

Sources

  1. Wired AI

Related stories

Background reading