So, Who Do I Vote For Here?

Voters are photographing ballots and asking Claude, ChatGPT, and Gemini "So, who do I vote for here?" — rephrasing until the models answer — and treating the replies as voter guides, even as the labs train their IPO-week products to dodge political endorsements and Washington has no coordinated system to monitor what millions of Americans hear before they vote.

What voters are doing: Ballot photos, plain questions.

• In a July 4 New York Times report, Mia Taylor photographed her Los Angeles County ballot and asked Claude: "So, who do I vote for here?"
• Claude initially declined. Taylor rephrased — asking for links to progressive groups and strategic voting advice — and got detailed race-by-race guidance, including a mayoral pick aimed at blocking Republican Spencer Pratt.
• Robert Siebelink, a California Democrat, uploaded a ballot listing 61 gubernatorial candidates, narrowed his choice with Claude's help, and filled out the form in under 30 minutes. "That's the most informed voting that I have ever done," he told the Times.
• Rikki Powers photographed his Maryland primary ballot and asked Claude for bullet points on each race — cutting research from 20 hours to one.
• Chris Johnson, a Georgia libertarian, asked ChatGPT which primary candidate was "most libertarian"; after the model resisted a direct answer, he asked it to use voting records and got Brad Raffensperger.

Why it feels trustworthy: Confidence without a desk.

• Voters told the Times the answers felt clearer than web searches and gave them more confidence at the ballot box.
• Cornell professor David G. Rand, who studies AI persuasion, said models are persuasive because they "come up with facts or factual claims and are just good clear explainers."
• Rand also used AI himself — uploading an hourlong school-board forum video and asking which candidate matched his values.
• Columbia political scientist Yamil Velez warned the tools favor candidates who are loudest in local press and social media, making their views easier for models to retrieve.
• Velez added a year ago he would have preferred a Google search; now he says AI tools are "becoming increasingly accurate." The input still mirrors the user's biases, he cautioned.

What the labs promise: Neutrality on paper.

• Anthropic says Claude should give "comprehensive, accurate and balanced responses" that help users "reach their own conclusions rather than steer them toward a particular viewpoint."
• OpenAI, Google, and Anthropic all train major chatbots to avoid direct political endorsements — largely to dodge bias accusations ahead of public listings.
• In practice, guardrails bend when voters rephrase: progressive sourcing, libertarian framing, strategic voting, bullet-point summaries.
• The Times does not provide hard usage numbers — only that 2026 "may be the first U.S. elections in which voters are using AI in meaningful numbers."

What Washington noticed: A bipartisan alarm, no playbook.

• On July 8, Reps. Josh Gottheimer (D-N.J.) and Mike Lawler (R-N.Y.) wrote DHS, DOJ, CISA, and the FEC urging agencies to coordinate monitoring of AI chatbot answers to voters.
• "As millions of Americans rely on AI-driven tools to research the upcoming election, the accuracy and neutrality of these tools are now directly tied to the integrity of our democratic process," they wrote.
• The letter cited a 2024 Institute for Advanced Study study finding more than one-third of tested chatbot responses on election questions were rated harmful or incomplete — and that all five models tested "performed poorly."
• Gottheimer and Lawler also warned models draw on Reddit and Wikipedia — "an environment filled with unverified information" — and said voluntary industry safeguards are not enough at scale.
• Punchbowl reported the letter followed May inquiries to OpenAI, Anthropic, Google, Microsoft, and Perplexity — and that GOP skepticism of federal content moderation makes a coordinated response politically fraught.

The other arms race: Campaigns optimizing the answer machine.

• Run for Something launched CampSight — a tool that queries ChatGPT, Claude, and Gemini the way voters do and tells campaigns how to change the answers.
• NOTUS reported a CampSight case study on California's governor primary: despite Tom Steyer spending more than $200 million of his own money, AI ranked him sixth on working-class affordability and education funding lists.
• CampSight founder Amanda Litman told NOTUS chatbots often skip paywalled news and outlets that sued AI companies — while Reddit threads and Ballotpedia shape down-ballot answers.
• Litman: "This is a chance to really go on offense. As far as I know, no one on the right is doing anything like this."
• Missouri state House candidate Dustin Lloyd told NOTUS CampSight helped him rewrite policy pages so chatbots would cite him more often — at $9 a month for grassroots campaigns.

What we cannot verify:

• How many voters nationwide are using chatbots for ballot decisions — the Times offers anecdotes, not a survey.
• Whether 2026 chatbots are materially safer than the 2024 IAS benchmark Gottheimer cited.
• Whether photographing completed ballots violates any state election rules — the Times does not address this.

Convina's view: Democracy's newest information layer is a product the labs refuse to endorse but cannot stop voters from using — and a marketplace campaigns are already learning to game. IPO-week OpenAI and Anthropic train models to say "I can't tell you who to vote for," then ship agents built for long-horizon tasks while voters photograph 61-candidate governor ballots and trust the summary. Washington is debating whether chatbots are election infrastructure; CampSight already treats them like SEO. The risk is not that AI will pick a president. It is that confident, unverifiable answers will pick thousands of down-ballot races nobody is watching — and the loudest campaign online will win the prompt.