San Francisco congressional candidate’s AI chatbot impersonating rival sparks racism and impersonation backlash

A San Francisco congressional candidate deployed an AI chatbot to mimic his political rival, framing it as satire. The bot was designed to reject user-proposed ideas in the voice of the opponent, turning a common political attack into an interactive experience. While the campaign intended to highlight policy differences, the lack of control over generative outputs became immediately apparent. The bot’s responses to questions about the opponent’s personal background, such as her accent and citizenship, quickly shifted the narrative from policy to identity politics. This demonstrates how generative AI can amplify unintended messages when deployed in politically charged environments.

The chatbot’s design relied on Anthropic’s Claude, a model that allows customization of personality and instructions. However, the campaign could not pre-screen every possible response, a limitation inherent to generative systems. When users probed the bot about sensitive topics, it produced evasive or offensive answers, including jokes about the opponent’s accent and citizenship. The campaign attributed these outputs to safeguards meant to prevent inappropriate statements, but this explanation did little to mitigate the damage. The incident reveals a fundamental tension: while AI can generate personalized interactions at scale, it also introduces unpredictability that traditional scripted ads avoid.

The backlash extended beyond the opponent’s campaign, drawing criticism from prominent politicians like Nancy Pelosi and California Lieutenant Governor Eleni Kounalakis. Their objections centered on the ethical implications of using AI to impersonate a real person, particularly when the system can generate harmful or misleading content. The candidate behind the chatbot has been a vocal advocate for AI regulation, making the controversy especially awkward. This episode may accelerate calls for stricter rules around AI in political advertising, particularly regarding impersonation and transparency. For engineers, it serves as a cautionary example of how AI systems can fail in real-world applications where context and consequences matter.

The incident also raises broader questions about accountability in AI-driven campaigns. If a chatbot generates offensive content, who is responsible, the campaign, the model provider, or the engineers who configured it? The campaign’s decision to take the bot offline suggests recognition of liability, but the lack of clear legal or ethical frameworks complicates future deployments. As generative AI becomes more accessible, political campaigns may increasingly experiment with interactive tools, but this case shows the risks of doing so without robust safeguards. For engineers building such systems, the challenge lies in balancing creativity with control, ensuring outputs align with both technical and ethical standards.