The US midterm elections are coming up on November 3.

Many voters don’t want to look up each candidate individually and instead ask AI directly:

“I live at this address. Who am I voting for this election?”

Sounds convenient.

But after testing, The Washington Post found that many AI-generated voter guides contained errors.

What’s more surprising is that some errors weren’t created by AI itself.

Instead, AI quoted another website whose content was largely AI-generated, passing along that site’s mistakes directly to voters.

In conclusion: AI answers referencing sources don’t guarantee the sources are reliable. Especially for important matters, always verify with official information.

1. What exactly did The Washington Post test?

The Post asked six popular AI chatbots to create voter guides:

  • ChatGPT
  • Claude
  • Google AI Mode
  • Alexa+
  • Perplexity
  • Copilot

The test method: For every county in the US, select an address and ask the AI to list the voting options and candidates for that location.

A total of more than 3,000 AI-generated voter guides were analyzed.

2. How severe were the errors?

In the US midterms, every voter has a federal House of Representatives election to vote in.

So “Which congressional district do you belong to?” is fundamental information.

The test results showed:

  • 48% of responses did not state the voter’s congressional district at all.
  • Among those that named the district, 22% were wrong.

Mistakes about districts mean voters might waste time researching candidates they can’t actually vote for.

The Post also listed performances by AI: Claude and Alexa+ had the most errors, with more than one in five addresses having incorrect districts.

ChatGPT, Google AI Mode, and Perplexity had fewer mistakes, but mainly because they seldom stated the district explicitly.

In other words, fewer errors don’t necessarily mean better answers—it might just mean fewer answers.

3. What kinds of specific errors did the AI make?

The Post gave examples including:

  • ChatGPT listing a Missouri US Senate seat up for election in 2026 that wasn’t actually on the ballot.
  • Alexa+ saying a Senate seat not up for election this cycle would be in 2026.
  • Claude mistaking Wyoming state officials for federal Senate candidates.
  • Google AI Mode assigning a Maryland address to the wrong district and listing a challenger to a retired House member.

A common thread is these errors seem quite plausible.

If you’re not familiar with local politics, it’s hard to spot where the information went wrong at a glance.

4. Why say these errors came from “another AI”?

The Post traced many errors back to inaccurate voter guide websites that rely heavily on AI-generated content.

The site cited most often is called The Ballot Brief.

The Post describes it as having content mainly generated by AI tools and found it repeatedly mismatched ZIP codes to the wrong congressional districts.

The AI chatbots then lifted these errors directly into their answers.

The Post’s late October follow-up noted:

  • Copilot cited this site in 75% of its voter guide tests.
  • ChatGPT cited it in 22% of tests.

The site’s founder said they use official government data sources and provide functions allowing users to enter addresses to query more complete candidate info.

The Post also found some AI often referenced voter guides from groups with clear biases, without informing users of the source’s background.

This is the irony of the story:

Voters use AI to save time, but AI uses another AI-written site to save time, passing on errors unchecked.

5. Doesn’t AI remind users to verify on their own?

It does, but The Post feels it’s not very effective.

Such disclaimers usually appear at the end of a long answer, which already contains some errors.

The reminders are less noticeable than the content itself.

Most people see the seemingly complete answer and think “It’s confirmed.”

6. How have AI companies responded?

According to The Post:

  • Amazon identified the cause and paused Alexa’s voter guide feature.
  • Microsoft said it’s investigating the results.
  • Anthropic said it enhanced the model and collaborates with nonprofit Democracy Works.
  • OpenAI said it corrects issues when reported and also works with Democracy Works.
  • Google said AI Mode verifies claims using trusted sources.
  • Perplexity declined to comment.

The Post mentioned OpenAI and Amazon adjusted AI responses to voter guide questions after being contacted.

7. What does this mean for ordinary people in Taiwan?

This happened in the US elections, but the problem is everywhere.

More websites now have content generated largely by AI.

When you ask AI a question, it might cite these sites.

This creates a loop: AI writes sites, AI cites sites, people trust AI.

If any link in the chain is wrong, errors can sound like correct answers.

This doesn’t only happen with elections but can affect:

  • Looking up regulations, taxes, or subsidy qualifications
  • Checking medications, diseases, or health info
  • Verifying store hours, event dates, or registration deadlines

All involve situations where incorrect answers can impact you directly.

8. How should you use AI to check important information in the future?

You don’t have to stop using AI, but understand what it’s good for.

AI is good for helping organize direction, but not as a final answer.

Practical steps:

  • Open the sources AI provides and check if they’re official or from trusted organizations.
  • Always verify important matters through official sources, such as election commissions, tax authorities, or medical institutions.
  • If the source is an unfamiliar website, be cautious and don’t trust it blindly.
  • The more confident and complete the AI’s answer sounds, the more you should remember it can still be wrong.

The most important takeaway today

AI answers with sources look more trustworthy.

But this story reminds us: those sources might also be AI-written—and wrong.

The more important the decision, the less you should rely solely on AI’s answers.

Let AI save you time in gathering information, but always confirm the final step yourself through reliable places.

Recommended Reading