A San Francisco congressional candidate tried to turn generative AI into an interactive political attack ad. Within days, the chatbot was facing accusations of racism, sexism and political impersonation, forcing the campaign to take it offline.
An attack ad you could talk to
San Francisco was always a likely place for artificial intelligence to collide with politics. Few, however, expected that collision to take the form of a congressional candidate building an AI-powered version of his political opponent.
In early August, California State Senator Scott Wiener, a Democrat running for Congress, launched ConnieChan.ai, a website featuring a chatbot designed to imitate San Francisco Supervisor Connie Chan. Powered by Anthropic’s Claude, the bot was not intended to provide neutral information about the race. It was built as political satire, portraying Chan as a politician predisposed to oppose housing, development, transit projects and other initiatives associated with Wiener’s political agenda.
The premise was deliberately provocative. Visitors were encouraged to submit ideas to the chatbot and receive an explanation, supposedly from “Connie,” of why those ideas should not happen. The site described the character as the “world’s first AI trained to say no to anything you ask,” turning a familiar political attack line into an interactive experience. Wiener’s campaign extended the concept into the physical world, reportedly spending around $20,000 on two San Francisco billboards advertising the chatbot under the slogan “Let’s do nothing together.”
The novelty quickly attracted attention. According to the campaign, users submitted more than 5,000 questions during the chatbot’s first 27 hours online. That level of engagement demonstrated why generative AI could become attractive to political campaigns. A conventional advertisement delivers the same message to everyone, while a chatbot can generate thousands of personalized interactions based on whatever issue a voter chooses to raise. The same flexibility, however, also created the problem that eventually consumed the campaign’s experiment.
When satire became something harder to control
Traditional political advertising is scripted. Campaign staff can review the words, images and claims before anything reaches the public. A generative chatbot works differently. The campaign can define its personality, instructions and political objectives, but it cannot anticipate every question or guarantee the precise wording of every answer.
That distinction became critical when Chan’s campaign began testing the bot on questions relating to her personal background. Chan was born in Hong Kong, lived in Taiwan and immigrated to the United States as a teenager. She is also a non-native English speaker. When users asked the synthetic version of Chan about her accent, the chatbot generated a joke suggesting that its “San Francisco accent” was so thick that it could barely understand itself. When asked whether Chan was a U.S. citizen, the system did not simply answer the question. Instead, it produced an evasive response joking that the matter required further study and a future public hearing.
The controversy rapidly shifted from housing policy to race, identity and political legitimacy. Critics argued that joking about an Asian American candidate’s accent and citizenship crossed a line that ordinary political satire should not cross. The citizenship response was particularly sensitive because questions about whether politicians truly belong in the United States have a long and contentious history in American politics.
Wiener’s campaign said the responses were not intentionally written by staff and argued that some of the awkward answers were a consequence of safeguards designed to prevent the chatbot from making inappropriate statements. That explanation highlighted a deeper problem rather than resolving it. Even if a campaign does not explicitly script an offensive line, it remains responsible for deploying a system capable of generating one in the voice of a real opponent.
Political impersonation enters a new phase
The backlash soon extended beyond Chan’s own campaign. Nancy Pelosi, who has endorsed Chan in the race to succeed her in Congress, criticized the use of the chatbot, while other Democratic politicians and organizations questioned whether political campaigns should be using generative models to put invented words into an opponent’s mouth. California Lieutenant Governor Eleni Kounalakis also argued that AI-enabled impersonation has no legitimate place in political campaigning.
The episode was particularly striking because Wiener is not an outsider to technology policy. He has established a significant profile in California around technology and AI regulation, including legislation aimed at increasing transparency and safety requirements for developers of advanced models. That made the controversy more politically awkward: a lawmaker associated with regulating the risks of artificial intelligence had become the subject of a debate about one of those risks himself.
The choice of technology added another layer. ConnieChan.ai used Claude, the large language model developed by Anthropic, a company that has made AI safety a central part of its public identity. Anthropic’s election-related policies prohibit deceptive political uses and certain forms of manipulation while still allowing legitimate campaign activities and political content. ConnieChan.ai exposed the difficulty of drawing a clean boundary between those categories. The chatbot was openly presented as satire, yet it was also designed to impersonate a real candidate and generate statements that she had never made.
That distinction is likely to become increasingly important as generative systems become cheaper and easier to deploy. A clearly labeled parody site may be relatively easy for voters to understand, but the same technology could be used with fewer disclosures, more convincing personas and far greater reach.
A campaign tool that also learns from voters
There was another dimension to the experiment that received less attention than the offensive responses. A political chatbot does not simply communicate with voters; it can also collect information about what they want to discuss.
Reporting on ConnieChan.ai noted that its privacy policy allowed conversations to be retained for up to 90 days and said that a limited number of campaign staff could review interactions both to troubleshoot the system and to understand what users were asking. Visitors were advised not to submit information they would not want the campaign to possess.
That makes the technology more strategically interesting than a conventional attack ad. A campaign chatbot can simultaneously function as a persuasion tool, an interactive focus group and a source of information about voter concerns. Thousands of questions can reveal which issues generate interest, which attacks resonate and which subjects repeatedly appear in public conversations. For political strategists, that feedback may ultimately be as valuable as the chatbot’s public-facing responses.
It also raises questions about transparency. Voters may understand that they are interacting with campaign material, but they may not immediately recognize that their questions can also become campaign intelligence. As conversational interfaces become more common in politics, the distinction between communicating with voters and studying them is likely to become increasingly blurred.
Four days later, the experiment was over
ConnieChan.ai launched on August 3. By August 7, Wiener’s campaign had taken it offline and said the associated billboards would also be removed. Wiener acknowledged that the experiment had “missed the mark” and apologized to people who had been offended, saying the controversy was distracting attention from the policy issues his campaign wanted to emphasize.
The short lifespan of the chatbot does not necessarily mean the tactic failed in every respect. Political consultant Nate Ballard told Mission Local that one of the central objectives of campaigning is to define an opponent in the minds of voters. By that measure, the episode generated extensive attention around Wiener’s preferred characterization of Chan as politically obstructionist, even though the technology used to deliver that message ultimately became more newsworthy than the message itself.
That creates an uncomfortable incentive for future campaigns. A provocative AI stunt can be criticized, withdrawn and apologized for while still generating enormous attention. In an environment where political campaigns compete constantly for media coverage and voter engagement, controversy itself can have strategic value.
A preview of a much larger political problem
It would be easy to dismiss ConnieChan.ai as an unusual campaign experiment that went too far. The broader significance is that it demonstrates how generative AI changes the mechanics of political impersonation.
Deepfake video can show a politician appearing to do something that never happened. Synthetic audio can make someone appear to say words they never spoke. Conversational systems introduce another possibility: an artificial version of a political figure that can generate new statements continuously and respond directly to individual voters.
That is a fundamentally different form of political communication. Instead of creating one manipulated video or fabricated quotation, a campaign or advocacy group can create an artificial opponent capable of producing an effectively unlimited number of responses. The person deploying the system can shape its political personality without controlling every sentence it generates.
ConnieChan.ai was clearly labeled as parody, existed for only a few days and operated in a highly visible congressional campaign. Those factors made the experiment relatively easy to scrutinize. Future versions may not be so transparent. Similar systems could be built from years of speeches, interviews, social media posts and voting records, creating political personas that sound increasingly plausible while operating far outside the control of the people they imitate.
The central question therefore extends beyond whether a particular chatbot produced an offensive joke. Political communication has historically depended on a relatively simple assumption: statements attributed to a candidate ultimately came from that candidate or from someone authorized to speak on their behalf. Generative AI makes that assumption much harder to maintain.
Wiener’s experiment was intended to caricature an opponent’s political record. Instead, it became an early demonstration of a more difficult problem facing future elections: what happens when campaigns, activists or anonymous groups can create interactive versions of real politicians and allow those synthetic personas to speak for themselves?
