Voters turn to ChatGPT for ballot research
NPR interviews show how voters compare candidates—and why a researcher warns that answers may reinforce existing views.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:003:03
Read transcript
Voters are asking chatbots to help them understand their ballots—and a researcher warns that feeling informed may not mean learning more. NPR’s October 5 interviews found people using ChatGPT, Gemini and Claude to compare candidates, look into local races and check viral claims. Some said shrinking local-news coverage left them looking elsewhere for down-ballot information. In Morgantown, West Virginia, voter Adam Johnson asked ChatGPT to assess candidates against his political principles. It said it could research his criteria, but not tell him how to vote. On immigration, it gave positions for Senator Shelley Moore Capito, while noting that it found less detailed public information about Rachel Fetty Anderson. It declined to fill that gap with generic assumptions about Democrats. Another voter, Lisa Veldran, used Gemini to make a candidate-comparison table. The concern is subtler than invented facts. Researcher Rafael Batista told NPR a chatbot may choose information that supports the user’s existing preference, leaving them more confident without necessarily improving their understanding. He also said limited disclosure about commercial models’ training and source selection makes their neutrality hard to assess. OpenAI pointed NPR to its election safeguards, saying it monitors political bias and designs ChatGPT not to manipulate users. Those are the company’s stated safeguards, not proof that every exchange is neutral. Some users try to add checks. Lakshmi Iyer used Claude to build Ballot Lookup, a volunteer-built Pennsylvania candidate-information site, asking it That question of what a safeguard can really prove also comes up in OpenAI’s new text watermark. The company is starting a phased rollout for eligible ChatGPT and Codex text in the European Union; API developers worldwide can opt in for select models. The detector is initially limited to approved researchers and expert organizations. In OpenAI’s tests, replacing just ten percent of words with synonyms cut detection from about 92 percent to 66 percent. A positive result does not establish who authored a passage, whether it is-` The limits of a signal matter when software is used to guide decisions, too. A proposed nationwide class action alleges McDonald’s used AI and nonpublic business data to coordinate prices across franchise and company-owned restaurants. The claims are unproven. Reuters reported that the pricing system analyzes millions of daily transactions across nearly 14,000 U.S. locations. McDonald’s rejects the allegations, saying franchisees make their own choices and can decline recommendations. The dispute is whether shared advice can, And in a more everyday setting, Instinct is testing how an agent can act for a group without exposing everyone’s private account. Early-access users can invite friends into group chats, even if those friends have no Instinct account. Founder Noah Shinn says the shared agent cannot access participants’ personal accounts; personal agents must ask permission before sharing information or taking action. Pending replies pause if someone new joins. Broader access is promised “soon,” with no date. Across these stories, watch the gap
