Using generative AI for election-related research

概览

The episode examines how artificial intelligence is shaping the 2026 U.S. midterm elections—as a campaign issue, a source of political funding, a tool for producing advertisements, and a way for voters to navigate complex ballot information.

Marketplace correspondent David Brancaccio tests ChatGPT, Claude, and Google Gemini on a California ballot proposition. The experiment shows that chatbots can surface useful information, but only when users ask precise questions, compare multiple systems, and independently verify their sources.

The discussion also highlights AI’s persuasive power. Cornell researcher David Rand says a six-minute conversation with a purpose-built chatbot meaningfully changed some voters’ presidential preferences in a 2024 study, with effects several times larger than those of a standard political advertisement.

分段落总结

[01:16] The 2026 Midterms as a Major AI Election

[事实] AI has become an election issue, a source of campaign fundraising, a tool for creating political advertisements, and a research aid for voters.

[事实] California voters face 14 ballot propositions, creating a substantial amount of information to evaluate.

[推测] The episode frames 2026 as a turning point when generative AI becomes integrated into both political campaigning and individual voting decisions.

[01:50] Testing Three Chatbots on a California Proposition

[事实] Brancaccio asked ChatGPT, Claude, and Google Gemini about a California proposition concerning money in elections.

[事实] ChatGPT initially confused the current Proposition 4 with a different measure from 2024 and incorrectly said he could not vote on it.

[事实] Adding the exact proposition number, election date, and year produced more relevant answers.

[事实] Despite the initial error, the chatbots ultimately provided information that Brancaccio considered useful for making an Election Day decision.

[推测] The test suggests that underspecified prompts can cause chatbots to retrieve outdated or contextually incorrect election information.

[02:22] Hallucinations and Source Verification

[事实] A documented chatbot error in Scotland involved inventing a nonexistent election candidate.

[事实] Brancaccio says users must check the links supplied by chatbots rather than accepting their answers at face value.

[事实] Some cited sources were authoritative, including Ballotpedia and the California Secretary of State’s office, while others required additional investigation and context.

[事实] One source described itself as an independent newspaper, but closer examination revealed conservative branding.

[推测] Assessing a chatbot’s answer requires evaluating not only factual accuracy but also the authority and possible ideological perspective of its sources.

[03:13] Comparing Multiple AI Systems

[事实] Brancaccio recommends asking several chatbots the same question and comparing their answers.

[推测] Disagreement between systems can help alert voters to claims that need further verification, although agreement alone would not prove that a claim is correct.

[04:32] How Strongly Can AI Influence Voters?

[事实] Cornell professor David Rand discussed a special chatbot his team built before the 2024 U.S. presidential election.

[事实] An average conversation of about six minutes meaningfully changed some participants’ views about whether to vote for Kamala Harris or Donald Trump, with movement occurring in both directions.

[事实] Rand says the measured effects were approximately three to four times larger than those associated with watching a standard political advertisement.

[事实] Brancaccio notes that people may respond well to the large amount of policy information AI can produce quickly.

[推测] Conversational AI may be more persuasive than conventional campaign advertising because it can deliver extensive information through an interactive, individualized exchange.

[05:36] Practical Advice for Voters Using Chatbots

[事实] Users should make their election-related questions specific.

[事实] They should consult multiple chatbots, compare the results, and inspect the links behind the answers.

[事实] Source quality varies: some links may be highly useful, while others may be unreliable.

[事实] Chatbots can behave like “people pleasers” by telling users what they appear to want to hear.

[事实] Brancaccio suggests explicitly instructing a chatbot to give a direct answer rather than reinforcing the user’s preferences.

[推测] Responsible use of AI for election research requires active skepticism and deliberate prompting, not passive acceptance of generated summaries.

播客点评/总结

The episode offers a concise and practical introduction to AI’s dual role in elections: it can help voters process complicated information, but it can also provide outdated claims, fabricate facts, obscure source bias, and influence political preferences.

Its strongest feature is the combination of a real-world chatbot test, concrete verification advice, and research on political persuasion. The recommendations—specify the election and date, compare systems, inspect sources, and challenge agreeable answers—are immediately useful.

[推测] The short format limits deeper discussion of the cited study’s methodology, the durability of its persuasion effects, and differences among commercial and purpose-built chatbots. The episode is best suited to voters, journalists, educators, and anyone using generative AI to research candidates or ballot measures.