Evaluating Text-to-Image Platforms' Content Moderation During the 2024 US Presidential Election

How do generative AI platforms’ content moderation policies handle the creation of political deepfakes? We evaluate how AI platforms mitigate this risk using an automated pipeline for politically diverse, externally valid evaluations of text-to-image (T2I) systems during the 2024 US Presidential election. Our system transformed media references to candidates into prompts for generative AI systems and sent prompts to three prominent T2I platforms each week for the final three months of the 2024 campaign. We first show that the platforms took fundamentally different approaches to content moderation, with little consistency in blocking behavior between platforms. We then show that there is little consistency in the blocking behavior within platforms over time. Almost no prompts were blocked in every week of our collection and Stability AI allowed almost all prompts featuring political figures until a sudden change two weeks before the 2024 election. Our findings highlight the importance of developing scalable context-specific approaches to monitoring T2I platforms.