Google Ads has introduced a new artificial intelligence system designed to significantly improve the detection of fraudulent advertisers and policy violations across its advertising platform. The system, based on a newly developed large-scale AI model, is already active and has demonstrated a major leap in performance compared to previous detection methods.
According to Google’s internal research, the updated model has increased fraud detection rates by more than 40 percentage points in key areas, while maintaining extremely high accuracy levels. In some policy categories, the system reportedly achieves close to perfect precision, greatly reducing the risk of false positives that could impact legitimate advertisers.
A New Foundation Model for Advertiser Analysis
The new AI system is known internally as ALF, short for Advertiser Large Foundation Model. It represents a shift toward foundation-style models that can process and combine multiple types of information at once. Rather than relying on isolated signals, ALF evaluates advertiser behavior using a broad range of inputs, including ad text, images, video creatives, account history, billing patterns, and long-term performance data.
While none of these signals alone necessarily indicate wrongdoing, analyzing them together allows the system to build a more accurate picture of intent. This multi-dimensional approach enables the AI to distinguish between normal advertising behavior and coordinated attempts to exploit the platform.
Why Earlier Systems Fell Short
Google’s research highlights several limitations in older fraud detection systems that ALF was specifically designed to address.
One major challenge involved the diversity and complexity of advertiser data. Advertiser profiles include both structured information, such as account age and payment methods, and unstructured content like images, videos, and landing pages. With hundreds or even thousands of data points per advertiser, traditional models struggled to interpret these high-dimensional datasets effectively.
Another issue was scale. Some advertisers maintain vast libraries of creative assets, making it easy to conceal a small number of malicious ads among thousands of legitimate ones. Earlier systems often failed to surface these hidden risks efficiently.
Reliability was also a concern. Fraud detection systems must generate confidence scores that businesses and regulators can trust. Excessive false positives can unfairly penalize honest advertisers, while overly cautious systems allow harmful actors to slip through. Previous approaches required frequent tuning to maintain acceptable performance levels.
Privacy-Centered Design
Despite relying on sensitive operational signals such as billing behavior and account metadata, the system was built with strict privacy protections in place. All personally identifiable information is removed before the AI processes the data. As a result, risk assessments are based on behavioral patterns and contextual signals rather than personal identity.
Identifying Suspicious Behavior Through Comparison
One of ALF’s most effective techniques involves comparing advertisers against one another at scale. Instead of evaluating each account in isolation, the model analyzes large groups of advertisers simultaneously. This allows it to establish a baseline of normal behavior across the ecosystem and more easily flag accounts that deviate in statistically meaningful ways.
By identifying these anomalies, the system becomes more effective at detecting coordinated fraud schemes and evolving tactics that might otherwise appear legitimate when viewed individually.
Performance Gains in Real-World Use
Testing results show that ALF consistently outperforms previous production systems across multiple benchmarks. The model delivers substantial improvements in both recall and precision under live conditions, not just in controlled test environments. While the larger model introduces slightly higher response times, performance remains within acceptable operational limits and can be optimized further using specialized hardware.
Despite the increased computational demands, the system is already handling millions of requests per day, demonstrating that the trade-off between speed and accuracy has been successfully managed.
Current Use and Future Potential
At present, ALF is integrated into the Google Ads safety infrastructure, where it is used to identify advertisers that violate platform policies. There is no indication that it has been extended to other Google products, such as Search or business listings.
However, Google’s research suggests broader potential applications. Future iterations may incorporate time-based analysis to better detect evolving fraud patterns. The underlying model could also support improvements in audience modeling and ad creative optimization, expanding its role beyond fraud prevention
