Google blocked 5.1 billion ads in 2024. That’s not a typo. Five billion ads that never made it to a single search result page, a YouTube pre-roll, or a Display Network banner — stopped before anyone had to see them. And according to Google’s 2025 Ads Safety Report, a significant chunk of that enforcement muscle now runs on Gemini. The question worth asking isn’t just how they did it — it’s whether AI-driven ad moderation is actually working, or whether the numbers are starting to obscure a more complicated picture underneath.
The Scale Problem Google Has Always Had
To understand why this matters, you have to appreciate the sheer size of what Google is trying to police. Google Ads serves billions of impressions every single day across Search, YouTube, Gmail, and the Display Network. Human reviewers were never going to cut it at that scale. The company has been layering in machine learning for ad review since at least 2017, but the shift to large language model-based enforcement — specifically Gemini — marks something qualitatively different.
The old systems were good at catching known patterns. A flagged URL. A banned keyword. A template that matched a previous scam. What they struggled with was novelty. Bad actors got smart. They’d rotate landing pages, swap out trigger words, use regional dialects, or bury deceptive claims three clicks deep into a site that looked perfectly legitimate on the surface. A rule-based classifier misses a lot of that. A model trained on billions of documents does not.
Google started integrating Gemini into its ads policy enforcement pipeline through 2024, and the 2025 report is the first time we’re seeing a full-year picture of what that looks like in practice. The results, at least by Google’s own accounting, are striking.
What Gemini Actually Does in the Ads Pipeline
Google hasn’t published a full technical breakdown — which is mildly frustrating — but from the report and surrounding documentation, the Gemini integration touches several distinct enforcement layers.
Understanding Ads at the Semantic Level
Traditional classifiers look at signals: URLs, ad copy keywords, image hashes. Gemini-powered review reads ad content more like a person would. It can understand that an ad claiming “doctors hate this one trick” is structured like a known scam even if none of the individual words are flagged. It catches implication, not just explicit statements. That matters enormously for categories like financial fraud, fake health products, and impersonation scams — all of which rely on being just vague enough to slide past keyword filters.
Landing Page Analysis at Depth
One of the bigger gaps in older enforcement was that review stopped at the ad itself. An advertiser could run a clean, compliant ad and then send users to a completely deceptive landing page. Gemini-based tools can now crawl and analyze destination pages with much more nuance — understanding context, claims, and whether the page materially matches what the ad promised. Google says this was a major contributor to taking down 1.3 billion advertiser accounts in 2024, up sharply from previous years.
Faster Detection of Emerging Scam Patterns
Here’s where the LLM advantage really shows. When a new scam type emerges — say, a wave of fake AI tool promotions or cryptocurrency giveaway schemes — it used to take time to update classifiers and retrain models. Gemini’s generalization capability means it can recognize a new scam as structurally similar to known ones without needing explicit retraining on that exact variant. That closes the window bad actors rely on.
The report’s key figures paint a clear picture of the scale:
- 5.1 billion ads blocked before they ever served to users
- 1.3 billion advertiser accounts suspended — the largest single-year account action Google has reported
- Over 1 billion ads restricted rather than fully removed (meaning limited to certain audiences or geographies)
- Significant increases in enforcement against financial fraud, deepfake celebrity scam ads, and impersonation of government entities
- New enforcement categories added for AI-generated deceptive content specifically — a first in Google’s policy framework
The Deepfake Problem Is Now Officially on Google’s Radar
Celebrity Impersonation and Synthetic Media
One area where Gemini-powered enforcement has been specifically called out is deepfake ads. These are ads using AI-generated video or audio of celebrities — Elon Musk promoting a fake crypto platform, a synthetic Taylor Swift selling dubious health supplements — that have been a persistent and genuinely harmful scam vector for the past two years.
Google says it added dedicated policy categories and detection capabilities for synthetic media deception in 2024. Gemini’s multimodal capabilities — the fact that it can process images and video, not just text — are central to this. Detecting a deepfake ad requires understanding visual and audio artifacts alongside the semantic context of the claim being made. That’s not something a text-only classifier handles well.
This is an arms race, and I wouldn’t be surprised if the next iteration of this report shows bad actors specifically engineering synthetic content to fool Gemini’s detectors. That’s just how this goes. But at least the capability is now in the pipeline.
Why the Numbers Alone Don’t Tell the Full Story
Here’s the thing: Google both creates the ad platform and enforces its rules. There’s an inherent tension there that no amount of impressive statistics fully resolves. When Google says it blocked 5.1 billion ads, we’re taking Google’s word for what counted as harmful versus legitimate. The company doesn’t publish false positive rates — how many legitimate advertisers got caught up in enforcement sweeps. Small business owners and independent publishers have been complaining about wrongful account suspensions for years, and AI-powered enforcement at scale almost certainly increases that problem even as it improves overall accuracy.
Independent researchers at organizations like the Global Disinformation Index have documented cases where harmful ads continued to run on Google properties even after being reported through official channels. The 5.1 billion figure is impressive. The question is what percentage of harmful ads that figure actually represents.
How This Fits the Broader AI Safety Push
Google’s move here is part of a wider industry pattern of deploying foundation models for trust and safety work that goes well beyond ad moderation. Microsoft has been integrating GPT-4-class models into its content safety APIs. Meta uses its own internal LLMs for moderation across Facebook and Instagram. The bet, across the board, is that general-purpose reasoning models outperform narrow classifiers for nuanced policy enforcement.
It’s a bet that’s largely paying off on precision — catching more sophisticated violations. Whether it’s paying off on recall — not missing violations — is harder to verify from the outside. For a deeper look at how AI companies are thinking about trusted access for security-sensitive applications, our piece on OpenAI GPT-5.4-Cyber and trusted access for defenders covers the parallel challenge in cybersecurity contexts.
What’s interesting about Google’s specific approach is the integration depth. This isn’t Gemini as a bolt-on checker that flags ads for human review. According to the report, Gemini-based tools are making enforcement decisions autonomously at scale — with human review reserved for appeals and edge cases. That’s a significant shift in how AI is being used, moving from assistant to primary decision-maker in a consequential process.
And Gemini’s expanding role across Google’s products makes this kind of integration increasingly natural. The same model family powering the Gemini desktop app and consumer-facing features is now doing enterprise-grade policy enforcement work in the background. That’s either reassuring evidence of the technology’s maturity or a reminder of how much is riding on a single model family’s judgment calls.
What This Means for Advertisers and Users
For regular users, the short version is: the ads you see are statistically less likely to be outright scams than they were three years ago. Financial fraud ads, fake government impersonation, and AI-generated celebrity scams are getting caught earlier. That’s genuinely good.
For legitimate advertisers, particularly smaller ones without dedicated compliance teams, the picture is more complicated. AI enforcement at this scale will produce false positives. If your account gets suspended, the appeals process is slow and often opaque. Google has historically been poor at explaining why a specific account was actioned, and there’s no indication that changes with Gemini in the loop — in fact, explainability may get harder, not easier, as the decisions are made by a model rather than a human applying a specific rule.
For the broader ad-tech industry, this is a signal that IAB standards and self-regulatory frameworks are probably going to need to catch up with what AI enforcement actually looks like in practice. The policy questions around automated ad moderation — who’s accountable, what recourse exists, how errors are corrected — are still being worked out in real time.
Frequently Asked Questions
What is Gemini’s specific role in blocking harmful ads?
Gemini acts as a semantic understanding layer in Google’s ad review pipeline, analyzing ad copy, landing pages, and visual content for deceptive patterns that rule-based systems would miss. It’s especially effective at catching novel scam structures and AI-generated synthetic media used in fraudulent ads.
Does this mean all harmful ads are now being caught?
No, and Google doesn’t claim that. The report documents enforcement actions taken, not overall coverage of all harmful ads. Independent researchers have documented cases where harmful ads continued to serve after being flagged, suggesting gaps remain even with Gemini-powered enforcement.
What types of harmful ads saw the biggest improvement in detection?
According to the report, financial fraud ads, deepfake celebrity impersonation ads, and government entity impersonation scams saw significant enforcement increases in 2024. Google also added a new policy category specifically for AI-generated deceptive content.
Could Gemini-based enforcement wrongly block legitimate advertisers?
Almost certainly, yes. AI enforcement at scale produces false positives — it’s statistically unavoidable. Google doesn’t publish data on wrongful suspensions, and the appeals process remains slow. Small advertisers are most vulnerable since they typically lack the resources to contest enforcement actions quickly.
The 2025 Ads Safety Report is a credible signal that Gemini-powered enforcement is moving the needle on ad fraud in measurable ways. The next frontier isn’t blocking more ads — it’s building the transparency and accountability infrastructure that makes autonomous AI enforcement trustworthy enough to rely on at this scale. Google hasn’t solved that part yet, and the pressure to do so is only going to increase as these systems make more consequential decisions with less human involvement.
