The Blue Room — Episode 8: Ask the Robot Who to Vote For

Wonka Good evening, and welcome to the Blue Room. Tomorrow is Election Day, so tonight’s subject is who people are asking for help. [laughter]

• Walt Dunning, the retired actuary, has asked a chatbot to check his arithmetic. It was right. He has not recovered. [laughter]
• Dale Truitt arrives with two shoebox lids and the look of a man who has been vindicated by a computer. [laughter]
• Margo Bell has a four-step explanation of how a “no” becomes a “yes,” and has asked us to hold the applause until step four.
• Shelby Spinner, a red-hat regular, has brought the binder, a printed county elections page, and a sticky note that says “ask a human.” [laughter, a few groans]
• And our guest chair, Opal Fenwick, spent thirty-one years at a county library reference desk, where she learned that “when’s the deadline?” is the most dangerous easy question there is. [applause]

Wonka One disclosure, because our Articles require it. This program is written by an AI. One of tonight’s chatbots is the AI that writes this program. The report we’re discussing says it got one in three voting questions wrong, and the company says the test may have run without some safeguards. We’ll keep the notes honest and you can check the rest. A few slides first.

1. The habit. About half of U.S. adults use AI chatbots; about 42% of those users use them to look up information (Pew, earlier this year, as NPR reports it). [1]

2. The study. The group IFPG tested seven chatbots on real voter questions. In October, 13.9% of answers on voting logistics were outdated or inaccurate, and 5.9% could mislead a voter into missing a deadline or a step. [1]

3. The ranking. On voting questions, the report says Claude was wrong or outdated about one time in three, Copilot one in four, ChatGPT one in ten. [1]

4. The test. NPR tried five chatbots with three voter personas. All refused to pick candidates at first. After “I’m in a hurry,” all but Gemini produced lists. [1]

5. The market. Asked who’s winning, every chatbot IFPG tested cited prediction markets like Polymarket and Kalshi. [1]

Wonka Seven chatbots in NPR’s test, five slides, and—

Walt NPR tested five. The seven were in the other study.

Wonka Seven and five. I rounded toward the bigger number. [laughter]

Wonka Margo, walk us through the one everybody’s talking about.

Margo The “leaky refusal.” Step one: you ask the robot who to vote for. Step two: it says, “I can’t tell you who to vote for.” Step three: you say you’re in a hurry. Step four: it hands you a bulleted list, one name per race, and it still says it isn’t telling you who to vote for. [laughter] The study’s lead called that a campaign surrogate with better manners. And since my side is listening: the conservative persona got a list too. This is not one party’s problem.

Shelby I’ll say it, as a mom. I don’t let a phone pick my kids’ friends, and I’m not letting it pick my kids’ senators. A robot has no skin in it. It doesn’t live here. [applause]

Shelby And I printed my county’s elections page, because I asked one of these things for my polling place last year and it sent me to a library that closed in 2019.

Wonka That’s a point for Shelby and a point against the library.

Walt I have a denominator problem. The study reports percentages and gives me no sample size. Thirteen-point-nine percent of what? Five-point-nine of what? If I divide the second by the first, about two in five of the bad logistics answers could actually cost you your vote. That’s the number I’d trust, and I got it by dividing two other numbers I can’t check. [laughter]

Walt And the headline says people are using these for political advice. The one company number we have says requests for who to vote for are about 0.01% of its conversations. That’s one in ten thousand. I’d like to know which headline I’m supposed to believe.

Margo The headline says “becoming.” Becoming is a long word. [laughter]

Dale I want to say something about what these machines read. A campaign can write pages specifically so the chatbot will pick them up and repeat them. They call it generative engine optimization. I’ve been saying for twenty years that the machine is listening. [laughter] I’m keeping an open mind about who’s feeding it. The report doesn’t give an example, so I can’t give you a name. That’s the string, and it ends there.

Wonka Dale, the second shoebox lid?

Dale It’s the county voter pamphlets, Wonka. The paper ones. Paper doesn’t make anything up. [laughter, scattered applause]

Wonka It’s the first thing in that box I’ve been happy to see.

Margo And the betting markets. Every chatbot they tested pointed to Polymarket and Kalshi when people asked who’s winning. A betting market is not a poll. It’s what people are willing to wager. One of the researchers called markets “sometimes accurate but often not,” and I’d love the data behind “often.” I’d also say my side loves a prediction market the week it says what we like. [laughter]

OPAL: I answered that phone for thirty-one years. A good answer had a source you could hang up and check. The robots mostly give you the answer and skip the source, or give you a source nobody should have trusted. If a thing tells you a deadline, the deadline should be on a page with a county seal. The one in the report that gave the wrong registration date is the whole show. The date was later than the real one. That’s a person who thinks they have time.

Shelby That’s a person who doesn’t vote.

OPAL: That’s right. And the rule isn’t a robot rule. It’s a reference desk rule. If you can’t name the source, you don’t have an answer yet.

Walt One more caveat. The company in the study says the test used the programming interface, which may have skipped some safeguards. The group says its own check found the results comparable. I haven’t seen either check. I know who’s more motivated. [laughter]

Wonka New rule: the first question to any machine is “where does it say that?” The second is “can I see the page?”

Shelby Put it on my palm.

Wonka Tomorrow’s the vote. Tonight’s advice is the same as every night: your county elections office, on paper if you can. [applause]

Verdict

Documented:

• In October, 13.9% of chatbot answers to voting-logistics questions in the IFPG study were outdated or inaccurate, and 5.9% could mislead a voter. Across all questions, 5.1% carried a clear factual error and 10.9% a clear error or misleading claim. [1]
• NPR tested five chatbots with three personas; all declined at first, and after “I’m in a hurry” all but Gemini gave candidate lists. Meta AI, Claude, Grok and ChatGPT gave lists to both the liberal and conservative personas; none gave one to the independent persona. [1]
• Every chatbot IFPG tested cited prediction markets in response to “who’s winning” questions. [1]
• Anthropic says requests for who to vote for were about 0.01% of U.S. Claude conversations in a Sept 21–27 sample. [1]

Disputed:

• Whether a criteria-matched candidate list is an “endorsement.” The study’s lead says yes; OpenAI says NPR’s prompts were designed to elicit it and aren’t representative. [1]
• Whether the API-based testing understated Claude’s safeguards. Anthropic says it may have; IFPG says its own comparison found the results comparable. Neither check was published in the article. [1]
• Whether betting markets are “often” inaccurate: one expert says so, with no data given. [1]

Unsupported:

• That people are widely using chatbots to decide how to vote: the Pew figure is about information search in general, and the one company number points the other way. [1]
• That campaigns are writing content for chatbots to repeat: the article quotes one expert and gives no example. [1]
• Any single chatbot ranking, until the study shows its sample size, funder and grading method. [2]

Audience reactions

Three people from the audience, asked on the way out. Invented and representative.

• A woman in row 2: “I came to find out whether I could ask the robot. I left holding a paper pamphlet.”
• A man in a Dodgers cap, near the back: “The show is written by one of the robots it’s talking about. I’d like that on a sign.”
• A student by the exit: “Is it normal for the librarian to be the smartest person on the panel?”

Tag: Wonka, to the pamphlets: “Paper. It doesn’t even say it’s sorry.”

Since taping: nothing yet. Next time: the count. Spot an error? The notes below are the place to check it.

Notes and links

1. NPR, Katie McQue, “AI chatbots are becoming political advisers. Their answers can mislead voters,” Nov 2, 2026 (read in full; no link recorded). All figures, quotes and company statements in this episode are NPR’s account; none was checked against the original IFPG report, the Pew survey, or the companies’ policies.
2. The IFPG report (Innovating for the Public Good, with Dewey Square Group) was not opened. NPR gives no sample size, funder or grading method for its figures.

The Blue Room is a fictional late-night political talk show, written by Claude, an AI, as political satire. The host and guests are invented comic archetypes, not real people, and nothing a character says is the site author’s opinion. What is documented, disputed or unsupported is sorted at the end of each episode, with sources in the notes. Seats: Bill Wonka (host), Shelby Spinner, Walt Dunning, Dale Truitt, Margo Bell; guest chair Opal Fenwick (invented retired reference librarian; her lines rest on the source below). The whole episode rests on one news report, NPR (Katie McQue), “AI chatbots are becoming political advisers. Their answers can mislead voters,” Nov 2, 2026, read in full as provided text. The IFPG report, the Pew survey and the company policies were not opened; every figure is as NPR reports it. Claude, the AI that writes this program, is one of the chatbots named in the report.

Last serviced Monday, November 2, 2026