Picture a nightclub with one bouncer and a queue stretching four blocks. He is fast, he is polite, and he has a decent eye for a fake ID. But the line never stops moving, and management pays him by the number of people he lets through. Some dodgy characters get inside. Not because the bouncer is corrupt, and not because nobody cares, but because the maths of the door were never going to work out.
That is roughly the shape of Google’s advertising problem, and it is a useful one to sit with if you are trying to understand how automated systems behave in the real world. I write about AI agents for people who do not build them, and ad moderation is one of the clearest examples you will find of what these systems are actually good at and where they quietly fall down.
Automation is a filter, not a judge
The stated reason bad ads slip through is a combination of human oversight gaps and automated systems that miss some violations. That phrasing sounds like corporate hedging, but it is also just an accurate description of how machine review works. An automated reviewer is a filter. It catches patterns it has seen before, at enormous speed, across a volume no team of humans could touch. What it does not do is exercise judgment about something genuinely new.
A scam ad that looks like ten thousand previous scam ads gets caught. A scam ad that looks like a legitimate small business, right up until the moment someone clicks through and loses money, is a much harder call. The system has no concept of intent. It has features, thresholds, and a decision boundary.
This is the single most useful thing to understand about AI agents in general, whether they are reviewing ads, screening job applications, or flagging suspicious transactions. They are pattern matchers operating at scale. Scale is the whole point, and it is also the whole problem, because every percentage point of error becomes an enormous absolute number when you multiply it by millions of decisions.
Complaints as a new kind of signal
The genuinely interesting development is what changed in 2026. Google expanded its ad review process so that user complaints can now trigger throttling on search ads, not just policy violations detected by the system itself.
Limited ad serving is not a new mechanic. It has existed on Display and YouTube for years as a throttle Google applies to advertisers it does not fully trust yet, typically new accounts or accounts it cannot verify. What is new is the trigger. Complaints from actual humans now feed into the decision.
Read that as an admission, and a sensible one. If your automated reviewer cannot reliably tell a scam from a shop, you add another sensor. Real users encountering a bad ad in the wild are, collectively, a very good detector of the thing your classifier missed. They just happen to be a slow and expensive one, because the detection only happens after the harm.
This pattern shows up across AI systems that operate at scale. The first line of defence is automated and fast. The second is human feedback, fed back in as a correction signal. Neither works alone. The automation misses novel cases, and the humans cannot possibly review everything upfront.
The uncomfortable incentive question
There is a less charitable reading, and it deserves an airing. Never attribute to malice that which is adequately explained by stupidity, as the saying goes. But you do have to ask whether removing every ad that generates money is fully aligned with the interests of a business that makes its money from ads.
I do not think this is a conspiracy. I think it is something more ordinary and harder to fix. When a system is imperfect, the direction of its imperfection tends to follow the incentives of whoever tuned it. A filter set slightly too permissive costs the platform very little and earns it revenue. A filter set slightly too strict blocks legitimate advertisers, who complain loudly and have account managers. Those two errors are not felt equally by the company making the tradeoff.
That asymmetry is worth watching in any automated system you interact with, not just this one. Ask which mistake is expensive for the operator, and you can usually predict which mistake the system makes more often.
What this means for you
Two practical takeaways. If you are a person who sees ads, complaining actually does something now in a way it did not before. Your report is a signal in a pipeline, not a message into a void.
If you are an advertiser, the same 2026 changes cut in both directions. Complaint-driven throttling means your account’s standing depends partly on how users react to your creative, not only on whether you technically comply with policy. That is a different kind of risk to manage.
And if you are simply trying to understand how AI systems work, this is a good case to hold onto. The failures are not exotic. They are the predictable result of pattern matching at volume, tuned by someone with skin in the game.
đź•’ Published: