Google used Gemini to police ads

Google says Gemini now sits inside one of its biggest policing systems: the ad business, where it blocked or removed 8.3 billion ads in 2025. (blog.google) Google said its Gemini-powered tools stopped more than 99% of policy-violating ads before they ran, suspended 24.9 million advertiser accounts, and blocked 602 million scam-related ads tied to 4 million accounts. (services.google.com) The company also said Gemini helped cut incorrect account suspensions by 80% and process four times as many user reports as in 2024. In the U.S., Google removed more than 1.7 billion ads and suspended 3.3 million advertiser accounts in 2025, TechCrunch reported, citing the company’s country data. (services.google.com) (techcrunch.com) Ad screening is the work of deciding, at huge scale, whether an ad is legitimate before users see it. Google said Gemini improved that process by moving from keyword matching to understanding intent, which it used to detect impersonation, payment scams, and coordinated fraud campaigns. (blog.google) (services.google.com) That matters because Google’s ad network is one of the largest distribution systems on the internet, and the same model family is now showing up in other parts of Google’s business. Reuters reported on April 16 that Google is negotiating with the Department of Defense over a deal that would let the Pentagon use Gemini in classified settings. (blog.google) (usnews.com) Reuters said the talks would allow the Pentagon to deploy Gemini for “all lawful uses,” according to The Information’s reporting, while Google has proposed limits against domestic mass surveillance and fully autonomous weapons. Google did not publicly announce a deal as of April 18. (usnews.com) Google is also tightening the commercial plumbing around Gemini and Vertex AI, its managed platform for running models inside corporate software. GitLab said on April 14 that Google Cloud customers can now run GitLab Duo Agent Platform on Vertex AI models they already use and count that spending toward existing Google Cloud commitments. (ir.gitlab.com) GitLab framed that integration around “governed” model access, meaning companies can centralize which models their coding agents call and how those calls are managed. That is the same enterprise pitch Google is making in ads and, reportedly, in government: one model family, multiple controlled environments. (ir.gitlab.com) (blog.google) The numbers in Google’s ad report show where the immediate test is happening. Before Gemini reaches classified networks or more corporate agents, Google is using it to decide which messages billions of people never see. (services.google.com)