Ever wonder how Google keeps inappropriate ads off your screen? It's a monumental task, given the sheer volume of ads they process daily. A new research paper reveals Google's secret weapon: a cutting-edge AI system that can spot bad ads *before* they even go live. Instead of relying on traditional methods of training AI with countless labeled examples of bad ads (a time-consuming and resource-intensive process), Google's researchers have developed a clever shortcut. They use large language models (LLMs), the same technology that powers ChatGPT, to generate detailed text descriptions of what constitutes a policy violation. These descriptions are then turned into “co-embeddings,” mathematical representations that capture the semantic relationship between images and text. When a new ad image comes in, the system checks its similarity to these policy-violation co-embeddings. If it matches closely with descriptions of banned content (like tobacco products or misleading imagery), it’s flagged for further review or outright blocked. This allows for “zero-shot” moderation – meaning the system can identify violations without ever having specifically seen them before. The system is also surprisingly agile. Updating policies is as simple as tweaking the text descriptions, no lengthy retraining required. Early tests show this new approach catches over 100% more bad ads than previous methods, maintaining a high level of accuracy. This clever use of LLMs and co-embeddings is a game-changer for online safety. It allows Google to stay one step ahead of bad actors while maintaining a vibrant and trustworthy ad ecosystem. While the research focuses on ads, this technology holds exciting potential for other areas of content moderation, promising a safer and cleaner online experience for all.