Google Ads Targets Fraud with AI

Google Ads has quietly introduced a new AI system aimed at improving the detection of policy breaches and malicious activity among advertisers.

The company released a research paper on 31 December 2025 detailing the new model, which represents a significant upgrade over previous systems. According to the report, the AI has increased detection rates by over 40 percentage points while achieving 99.8% precision on certain policy checks.

ALF: Advertiser Large Foundation Model
The new system, known as ALF (Advertiser Large Foundation Model), is a multimodal AI capable of analysing text, images, video, and additional account signals such as account age, billing details, and historical performance data.

Researchers note that individual signals may not necessarily indicate fraud, but assessing them collectively provides a clearer picture of advertiser intent and behaviour.

They explain: “A key challenge is accurately understanding advertiser behaviour, which is essential for matching ads to users and detecting policy violations. Addressing this requires a holistic approach, integrating structured account information, creative assets, and landing page content.”

For instance, an account might be new, contain ads for a well-known brand, and have a single declined payment. While each factor alone may appear harmless, together they can strongly suggest fraudulent activity.

Three Key Challenges Overcome
ALF tackles three main issues that previous detection systems struggled with:

Diverse, High-Dimensional Data – Advertiser information comes in multiple formats, including structured account details and unstructured creative assets. High-dimensional data, with hundreds or thousands of variables, posed challenges for conventional models.
Large Sets of Creative Assets – Some advertisers may include thousands of images or videos, hiding a small number of harmful items among many legitimate assets. Older systems often failed to detect these anomalies.
Reliability and Trustworthiness – The AI must generate dependable confidence scores to avoid falsely flagging legitimate advertisers, maintaining accuracy without constant retuning.

Privacy Safeguards
While ALF processes sensitive signals, such as billing information, all personally identifiable information (PII) is removed prior to analysis. This ensures the system evaluates behavioural patterns rather than personal data.

Detecting Outliers with Inter-Sample Attention
ALF uses a technique called “Inter-Sample Attention” to enhance its detection capabilities. By analysing groups of advertisers together, the AI identifies normal behaviour patterns across the ecosystem, making it easier to spot unusual or suspicious activity.

Performance in Real-World Tests
Testing has shown that ALF significantly surpasses existing production systems, delivering simultaneous gains in both precision and recall. One policy saw recall improve by over 40 percentage points, while precision reached 99.8% on another.

Although the model is larger and slightly slower than previous systems, latency remains within acceptable production limits. Optimisation using hardware accelerators is possible, and the AI is already handling millions of requests daily.

Deployment and Impact
ALF is now integrated into the Google Ads Safety system, actively identifying advertisers violating policy. Currently, it is not applied to Search or Google Business Profiles, but future applications may include monitoring temporal trends, optimising creative, or modelling audience behaviour.

By combining a wide range of data signals and advanced AI techniques, ALF represents a major step forward in protecting the Google Ads ecosystem from fraudulent activity and improving overall advertiser integrity.