Google's New Ads AI Is Quietly Deciding Who Gets Seen And Who Gets Banned

Google Ads AI model for fraud detection

The Advertiser Large Foundation Model (ALF) is a multimodal AI system for advertiser understanding and risk detection. It processes text, images, video, and structured account data within a single model. The research describes ALF as focusing on advertiser intent and behavior for fraud and policy enforcement.

ALF evaluates signals such as account age, billing information, creative assets, landing pages, and historical performance metrics. While individual signals may appear harmless on their own, the authors note that their combination can indicate elevated risk when it matches known fraud patterns.

Key reported production results include:

• A recall improvement of more than 40 percentage points on one critical Google Ads policy compared with a previous production system.
• Precision of 99.8% on another policy while also increasing recall relative to the earlier system.
• Live deployment serving millions of risk evaluation requests per day inside Google Ads.

The paper states that ALF outperforms a heavily tuned existing risk detection pipeline built on traditional machine learning models, including deep neural networks, model ensembles, gradient-boosted decision trees, and logistic regression with feature crosses. In production evaluation, ALF delivers simultaneous gains in both precision and recall.

Background and technical design

Google's researchers describe three main challenges in advertiser risk detection at scale. The first is handling heterogeneous, high-dimensional data that spans structured attributes and unstructured creative content. The second is managing unbounded sets of ad creatives, where a few harmful assets may be hidden among many compliant ones.

The third challenge is generating reliable confidence scores that limit false positives while remaining stable in live systems. Earlier approaches had difficulty meeting all these needs at once for Google Ads safety workflows. ALF is presented as a single foundation model intended to cover multiple policies and data types.

ALF jointly embeds advertiser account features, ad creatives, and landing page content into shared representations. The model processes large batches of advertisers rather than scoring each account in isolation, enabling a mechanism the authors call Inter-Sample Attention.

Inter-Sample Attention allows ALF to compare behavior patterns across advertisers within each batch. According to the paper, this helps the model learn what typical activity looks like in the advertising ecosystem, making it more accurate at detecting outliers that may indicate policy violations or fraudulent intent.

Privacy, deployment, and performance metrics

According to the paper, ALF operates with strict privacy safeguards inside Google's infrastructure. Personally identifiable information is removed from advertiser data before the model processes it. The system focuses on behavioral and content patterns instead of explicit personal details.

The authors report that ALF is deployed within the Google Ads Safety system to identify advertisers violating Google Ads policies. The paper does not describe deployment in other Google products or surfaces. Its documented scope centers on fraud detection and policy enforcement in advertising.

ALF has higher latency than the prior system because of its larger model size, according to the research. The authors state that this delay remains within acceptable limits for production use and can be reduced with hardware accelerators. They report that the accuracy gains justify the additional computational overhead in the current deployment.

The paper also highlights future research directions. These include modeling temporal dynamics of advertiser behavior and applying similar techniques to audience modeling and creative optimization. These topics are described as potential extensions rather than current production uses.

Source citations

Primary information in this report comes from Google's official research publication: ALF: Advertiser Large Foundation Model for Multi-Modal Advertiser Understanding. Additional deployment details reflect content summarized from publicly available coverage by Search Engine Journal.