External Reference — Archived 2026-04-20

ChatGPT and Claude Will Be a Force in Elections. Nobody Knows What to Do About It.

A Washington campaign press corps account of how unprepared political operatives are for the chatbots that voters already use to decide who to vote for

Alex Roarty · Published April 20, 2026 · notus.org
Archived Reference. This is a summary and annotation of an external news article by Alex Roarty, archived for Kyanos research. Original: notus.org. NOTUS is the Washington news publication of the Allbritton Journalism Institute. Roarty covers politics and campaigns for NOTUS and previously reported for McClatchy, Roll Call, and National Journal.

Why This Reference Matters

This is the clearest mainstream-press confirmation to date of the problem Kyanos exists to solve. A reporter who covers campaigns for a living interviewed more than a dozen political strategists in both parties and found the same thing across all of them: voters are already asking chatbots who to vote for, and the people running campaigns have almost no idea what the bots say or how to influence it. The piece is useful precisely because it is not written by a vendor. It is a newsroom describing the gap, in the voices of the operatives living inside it.

The Core Finding

Millions of people now use chatbots like Claude and ChatGPT for everyday questions, and operatives expect voters to add a new one this cycle: who should I vote for. Most of them, the reporting finds, have no clue what the answer will be or what they can do about an unflattering one. Six months out from a November election that will decide control of the House and Senate, some of the most influential information sources online remain, in the article's framing, a black box to both parties.

Central Finding

The effort to shape what chatbots say about candidates is real but nascent and isolated. As one liberal group leader put it, nobody has written the playbook yet, so most people simply are not doing it. A Republican operative compared the moment to a space race in which the first rocket has not launched.

The Scale Behind the Urgency

The article anchors the stakes in usage data: an Elon University poll found a majority of Americans, 52%, now use large language models, a share likely to keep rising as traditional search traffic declines. A former Harris campaign official quoted in the piece frames it well: chatbots will be more politically influential than the most AI-skeptical person thinks, and less than the most AI-friendly person thinks, and either way that still makes them important. This is the same premise behind the Effective Persuasion Cost framework: AI answers are now a voter-contact channel sitting alongside TV, mail, and canvassing, whether campaigns are measuring it or not.

What the Bots Actually Say

The reporting includes live examples of chatbots answering pointed political questions. Asked about a Democratic Senate nominee's record on crime, ChatGPT returned a framing that laid out the partisan attack and the defense in the same breath. Asked which 2028 contenders care most about affordability, Claude produced a detailed breakdown that spent the most time on one governor and surfaced the strongest critique against him. The pattern the article notices: the models appear to weight the candidates who have been most prominent in news coverage. This is exactly the affective-framing dynamic the Kyanos remediation work addresses. Factual errors can be corrected; the framing a model reaches for is harder to move and matters more.

The Levers Operatives Named

Lever 1
Clear candidate-site content
Strategists familiar with answer-engine optimization said a candidate's own site needs direct, plainly written information about background and positions. This is the controllable surface, and it is the one most campaigns underinvest in.
Lever 2
Feeding many platforms, not one
Campaign content has to reach a range of sources, because different models lean on different inputs. The article notes Google's Gemini draws more heavily on YouTube, while Grok leans more on X. A single channel does not cover the field.
Lever 3
Wikipedia, and its limits
Wikipedia is described as a key source the models trust, but one campaigns find increasingly difficult to edit directly. It is influence without control, which is its own strategic problem.
Lever 4
Measurement as a prerequisite
The article reports that two chief analytics officers from the 2024 Harris campaign built a tool to track what Gemini, Grok, and ChatGPT say about candidates by state, store the responses, and surface the sources the bots cite. The recognition driving it: this is the first cycle in which these answers carry real weight.

The Honest Caveat

The piece does not oversell. AI company officials say the models are built to be neutral and resistant to manipulation, drawing from public news, video, and other sources to produce accurate answers. The experts admit nobody fully knows the best way to shape what the platforms say, and even the methods that work may be hard for individual campaigns to implement when they are already stretched reaching voters by other means. This is a feature for the Kyanos position, not a problem: measurement, structured authoritative self-description, and compliant content development are the defensible levers precisely because raw manipulation is neither reliable nor advisable.

How to Use This Page

Best for the skeptic who wants confirmation from outside the AI-vendor world that the problem is real and unsolved. It pairs naturally with the Last Mile and Endorsement Economy papers: this article establishes the gap in a working reporter's words, and those papers lay out the framework and the levers for closing it.

Archived for Kyanos research use  ·  Original article: Alex Roarty, NOTUS, April 20, 2026  ·  All content and analysis is the author's own