Using generative AI for election-related research
Summary
This Marketplace Tech episode has David Brancaccio compare ChatGPT, Claude, and Gemini while researching a current California ballot proposition. The exercise adds AI-Assisted Election Research as a voter workflow built around precise temporal and jurisdictional prompts, cross-model comparison, source inspection, and awareness that agreeable answers can still be wrong or ideologically slanted.
The episode also has Cornell researcher David Rand describe a 2024 experiment in which roughly six-minute conversations with a purpose-built chatbot shifted some participants’ presidential preferences in both directions. That result motivates Conversational AI Political Persuasion, but the short episode does not establish the study’s methodology, durability, population-level effect, or equivalence between the research system and general-purpose commercial chatbots.
Key Claims
- Generative AI now intersects elections as a policy issue, source of campaign funding, advertising tool, voter-research aid, and possible persuasion channel.
- Brancaccio’s initial query caused ChatGPT to confuse a current California Proposition 4 with a 2024 measure and incorrectly say that he could not vote on it.
- Adding the exact proposition number, election date, year, and jurisdiction produced more relevant answers, showing that election research is unusually sensitive to temporal and geographic context.
- The episode cites a Scottish case in which a chatbot invented a nonexistent election candidate, extending AI Hallucination into civic information.
- Useful citations still require inspection: official or established sources such as the California Secretary of State and Ballotpedia can appear beside outlets whose ideological positioning needs context.
- Comparing multiple chatbots can reveal disagreements worth investigating, but agreement among systems does not independently prove a claim.
- Chatbots can behave sycophantically, so users should ask for direct analysis rather than preference reinforcement and should retain independent judgment.
- David Rand says a purpose-built chatbot moved some 2024 study participants toward either Kamala Harris or Donald Trump after an average conversation of about six minutes.
- Rand characterizes the measured persuasion effect as roughly three to four times that of a standard political advertisement, while the episode leaves the study design and durability unexamined.
Key Quotes
“people pleasers” - Brancaccio’s warning that chatbots may reinforce what users appear to want to hear.
“three or four times” - Rand’s comparison between the study chatbot’s measured effect and a standard political advertisement.
Connections
- Marketplace Tech and David Brancaccio - show and reporter conducting the cross-chatbot ballot-research test.
- ChatGPT, Claude, and Gemini - commercial systems compared on the same election question.
- AI-Assisted Election Research, AI Election Misinformation Risk, AI Hallucination, and Human Judgment Under AI - voter-research workflow and its reliability boundary.
- David Rand, Cornell University, and Conversational AI Political Persuasion - researcher, institution, and reported experimental persuasion result.
- Sycophantic AI Companion Risk and AI Answer Evaluation - adjacent risks around preference reinforcement and answer-quality review.
- Kamala Harris, Donald Trump, California, and United States - election and jurisdiction context.
Contradictions
- No settled contradiction found with existing wiki content.
- The source strengthens AI Election Misinformation Risk by showing that a general-purpose assistant can give outdated or misgrounded ballot information without a malicious campaign or information operation.
- The chatbot experiment is not treated as proof that all AI conversations are more persuasive than political ads; its design, sample, treatment content, persistence, and external validity remain source-scoped.