Google Ads is enhancing its defenses against fraudulent advertisers and policy violations with the quiet deployment of a powerful new AI model named ALF (Advertiser Large Foundation Model). A research paper, dated December 31, 2025, reveals that this advanced system is already deployed, marking a significant leap in Google's ability to detect malicious activity with substantial improvements in both precision and recall.

ALF: A Multimodal Approach to Ad Fraud Detection

The new AI model, ALF (Advertiser Large Foundation Model), whose details were published in the paper dated December 31, 2025, is a multimodal system designed to analyze a comprehensive array of data points. Unlike previous systems, ALF doesn't just look at isolated signals; it processes text, images, and video content from ads, alongside crucial account information such as age, billing details, and historical performance metrics. This holistic approach allows the AI to develop a deeper understanding of advertiser intent and behavior, crucial for identifying sophisticated fraud schemes.

Researchers behind ALF emphasize that individual data points might appear innocuous on their own. However, when combined and analyzed by ALF, these factors can paint a clear picture of fraudulent operations.

"A core challenge in this ecosystem is to accurately and efficiently understand advertiser intent and behavior. This understanding is critical for several key applications, including matching users with ads and identifying fraud and policy violations.

Addressing this challenge requires a holistic approach, processing diverse data types including structured account information (e.g., account age, billing details), multi-modal ad creative assets (text, images, videos), and landing page content.

For example, an advertiser might have a recently created account, have text and image ads for a well known large brand, and have had a credit card payment declined once. Although each element could exist innocently in isolation, the combination strongly suggests a fraudulent operation."

Overcoming Previous Limitations

ALF was specifically developed to address three key challenges that hindered earlier fraud detection systems:

  1. Heterogeneous and High-Dimensional Data: Advertiser data comes in various formats, from structured account details to unstructured creative assets like images and videos. Furthermore, each advertiser can have hundreds or thousands of associated data points, creating a high-dimensional mathematical representation that conventional models struggled to process effectively.

  2. Unbounded Sets of Creative Assets: Malicious advertisers often hide one or two fraudulent assets among thousands of legitimate ones. Previous systems were overwhelmed by the sheer volume of creative assets, making it difficult to spot the anomalies.

  3. Real-World Reliability and Trustworthiness: A robust fraud detection system must generate highly reliable confidence scores to avoid false positives that could unfairly impact legitimate advertisers. ALF is designed to operate consistently without requiring constant manual recalibration.

Prioritizing Privacy and Safety

Despite analyzing sensitive signals like billing history and account details, ALF is built with stringent privacy safeguards. Before any data is processed by the AI, all personally identifiable information (PII) is meticulously stripped away. This ensures that ALF's risk identification is based solely on behavioral patterns rather than sensitive personal data.

The "Inter-Sample Attention" Advantage

ALF employs a unique technique called "Inter-Sample Attention" to enhance its detection capabilities. Instead of evaluating each advertiser in isolation, the model analyzes "large advertiser batches," comparing interactions across multiple accounts. This allows ALF to learn what constitutes normal activity within the broader Google Ads ecosystem, making it exceptionally accurate at identifying suspicious outliers that deviate from typical behavior.

Superior Performance Benchmarks

Google's researchers report that ALF significantly outperforms existing production baselines. In real-world production environments, ALF has delivered substantial and simultaneous gains in both precision and recall. On one critical policy, it boosted recall by over 40 percentage points, while achieving an impressive 99.8% precision on another.

While ALF's larger model size results in slightly higher latency (the time taken to produce a response), this remains well within acceptable limits for Google's production environment. The trade-off is justified by its successful deployment, where ALF processes millions of requests daily, delivering substantially better fraud detection performance at scale.

Deployment and Future Potential

ALF is now fully deployed within the Google Ads Safety system, actively identifying advertisers who violate Google Ads policies. While currently focused on ad safety, researchers indicate future work could explore time-based factors ("temporal dynamics") to catch evolving fraud patterns. They also suggest potential applications in audience modeling and creative optimization.

For a deeper dive into the technology, you can read the original research paper:

ALF: Advertiser Large Foundation Model for Multi-Modal Advertiser Understanding