More than half of American adults now have regular access to a generative AI assistant, and a measurable fraction are using that access to answer the one question pollsters have tracked for decades: who should I vote for [1]. The 2026 midterms are, by the evidence of this spring, the first major American elections in which voters are consulting AI in meaningful numbers before casting ballots — photographing their sample ballots, uploading candidate questionnaires, and asking Claude, ChatGPT, and Gemini to render judgment [1].

The civic-technology framing that has dominated coverage of this trend misses the structural problem. AI platforms are not neutral conduits for voter information. They are systems trained on text corpora that overrepresent candidates with high media and social-media profiles, and they have no disclosure obligation regarding which candidates they recommend, how often, or on what grounds [3]. A voter in a contested state legislative race who asks an AI chatbot for guidance is getting an answer shaped by how much text about each candidate exists on the open web — which quietly tilts the field toward whoever has been loudest, longest, and best-resourced online [2].

Campaign strategists know this and have begun acting on it. Run for Something launched a tool called CampSight that runs real browser sessions with major AI assistants to test what they tell voters about specific candidates, compiling the responses to let campaigns adjust their messaging [2]. The adjustment, as documented by a Times of San Diego investigation, is taxonomic: advisers now counsel clients to publish materials in bullet-pointed, structured formats that AI systems prefer — a new SEO-for-chatbots arms race that proceeds entirely without disclosure requirements [4].

No platform is required to tell a voter that its recommendation on a local race is the product of training data, not a nonpartisan research process. No candidate is required to disclose that they restructured their website to optimize chatbot responses. No regulator has yet defined AI voter guidance as a political ad requiring a paid-for disclosure, a media service requiring a political-broadcasting license, or a voter-assistance tool requiring accessibility and accuracy standards [3].

The Washington Post's own testing found that major chatbots varied significantly in how they handled contested political claims, with some deferring to incumbents and others to challengers based on the composition of their training corpora rather than any deliberate editorial judgment [1]. The tools sound authoritative precisely because their outputs do not look like search results, with links the reader must evaluate — they look like considered answers.

The absence of a regulatory category is not an oversight. It is the natural state of a technology that arrived faster than the legislative calendar. What makes this midterm a data point rather than a crisis is that the game is only beginning: campaign consultants building chatbot-optimization practices are operating in a market without rules, and the first election cycle is always the one in which nobody has optimized enough to shift an outcome.

-- MAYA CALLOWAY, New York