Advanced AI Improves Enforcement
Our safety teams have been utilizing artificial intelligence (AI) powered by machine learning to enforce our policies on a large scale. This technology has enabled us to identify and prevent billions of harmful ads before they reach users. However, the existing machine learning models, while highly advanced, have traditionally required extensive training with hundreds of thousands, or even millions, of instances of inappropriate content.
The Latest Language models (LLMs), on the contrary, can quickly evaluate and understand content in large volumes, while also capturing subtle nuances within the material. These enhanced reasoning capabilities have led to more extensive and precise enforcement actions on our more intricate policies. For instance, our guidelines against Unreliable Financial Claims, encompassing ads promoting get-rich-quick schemes, face challenges from increasingly sophisticated bad actors. These malicious entities adjust their strategies and customize advertisements to deceive users with new financial products or services like investment recommendations or digital currencies.
While traditional machine learning models are adept at identifying policy violations, the rapid and constantly changing nature of financial markets sometimes makes it difficult to discern between legitimate and fraudulent services, hindering the quick scalability of our automated enforcement systems to counteract scams. LLMs excel in promptly recognizing emerging trends in financial services, pinpointing patterns of malicious actors exploiting these trends, and distinguishing between a genuine enterprise and a get-rich-quick scheme. This has empowered our teams to be more agile in combating a variety of emerging threats.
We are just scratching the surface in harnessing the potential of LLMs for ads safety. Gemini, Google’s most advanced AI model launched publicly last year, integrates sophisticated reasoning capabilities that we are thrilled to incorporate into our ads safety and enforcement endeavors.
