What if the scam ad you reported last month is still running because nobody at Google has a reason to care? Not a conspiracy. Not incompetence, exactly. Just a system where the cost of missing a bad ad is lower than the cost of reviewing every ad properly.
I review AI tools for a living, which means I spend a lot of time watching automated systems get graded on their best day instead of their average one. Google’s ad review stack is the largest deployment of that pattern anywhere. And it keeps letting garbage through.
Automation does the volume, humans do the exceptions
The explanation for dodgy ads is boring: automated review misses some violations, and human oversight doesn’t cover the gap. That’s it. There is no secret room where someone approves crypto scams. There is a pipeline that processes an unfathomable number of ads, catches most of the obvious stuff, and hands off a thin slice to people.
The developer behind atomic14 put the charitable read well, quoting the old line: “Never attribute to malice that which is adequately explained by stupidity.” Fair. But the same post raises the uncomfortable follow-up, asking whether it’s really in Google’s interest to remove ads that generate revenue. That question doesn’t require malice. It only requires indifference, which is cheaper and scales better.
Here is what I keep coming back to as someone who tests AI systems: an automated moderation layer is only as good as the incentive to tune it. A spam filter that lets through junk costs you inbox clutter. An ad filter that lets through junk pays the company running it. Those two systems will not converge on the same accuracy target, no matter how good the underlying models get.
The 2026 shift moves the burden onto you
In 2026, Google expanded its ad review process so that user complaints can trigger throttling on an ad. On paper that sounds like accountability arriving. Look at the mechanic and it reads differently.
Limited ad serving isn’t new. It has lived on Display and YouTube for years as a throttle Google applies to advertisers it doesn’t fully trust yet, typically new accounts or accounts it can’t verify. What changed is the trigger. Complaints now feed a system originally built for unverified advertisers.
Two things follow from that. First, users become unpaid moderators. The detection layer that should have caught the ad up front now leans on the person who got hit by it. Second, the enforcement action isn’t removal, it’s throttling. Reduced serving. The ad slows down rather than stopping. For a legitimate advertiser caught in a complaint pile-on, that’s a revenue problem with no clear appeal path. For an actual scammer, it’s a speed bump.
This is the pattern I flag constantly in AI tools that promise safety through feedback loops: the feedback loop is real, the enforcement teeth are optional, and the vendor gets credit for the former while quietly skipping the latter.
Advertisers are losing control at the same time
The other half of the story deserves attention because it cuts against the idea that Google is simply tightening standards. The July 2026 Google Ads terms of service changes around AI automation authorize behavior that used to be the advertiser’s call. If Final URL Expansion is enabled, Google may direct traffic to landing pages you did not designate for that campaign. The new terms allow it.
Sit with that alongside the complaint-based throttling. Your ad can be sent to pages you didn’t choose, and complaints about your ads can throttle your account. You carry the risk, the automation makes the decisions.
Targeting is moving the same direction. Google’s own rule for 2026 is that exact-match and phrase-match keywords can still trigger ads above or below an AI Overview, but only broad match and keywordless targeting are eligible for placement inside them. Precision loses. The system’s judgment wins.
What this tells you about AI moderation generally
Google’s ad problem is the clearest public example of a thing every AI agent buyer should internalize. Automated review is excellent at scale and mediocre at edges, and the edges are where harm lives. Adding a complaint channel does not fix the edges. It relocates them onto users and then measures success by how many reports came in, not how many bad actors left.
If you’re an advertiser, the practical moves are unglamorous. Check your Final URL Expansion settings and decide deliberately whether you want Google picking your landing pages. Assume complaint-driven throttling can hit you without a clean explanation. Keep your own records of campaign performance so you can spot a throttle when it starts.
If you’re a user, keep reporting the dodgy ads. Just don’t mistake the report button for enforcement. It’s a signal to a system that has never been strongly motivated to act on it, and that’s the part no model upgrade is going to solve.
🕒 Published: