22 Sep 2025 Search

A must-read from Rob Leathern for anyone trying to comprehend the surveillance advertising problem or ad reform in general: Rob’s Notes 29 - The AI Ads Cash Machine.

The advantage is the whole factory: systems that identify users and extract features in milliseconds; multi-stage ranking models that combine different types of predictions; experiment platforms that measure long-term effects across thousands of overlapping tests; attribution systems that connect ads to purchases days or weeks later; and inference engines that serve predictions from massive databases cached across data centers.

Read the whole thing, and you’ll get why regulators and legislators going on about trackers and data brokers might as well be talking about junk faxes or Flash supercookies. By the time we get to Delete Day in California (save the date, it’s August 1, 2026) it might get listed as a retrocomputing event. ICYMI, the open web is in rapid decline as an ad medium, according to Google, and they ought to know.

It’s the duopoly’s ad business now (Meta cut the old-fashioned data brokers out of the loop back when they were still Facebook Inc.) and most of the ad-related privacy harms that affect you today are happening many layers deep inside a Meta or Google ML system, not in places where you still have access to tools to research or block them. These ML systems are complex (deliberately so since that multi-year housing discrimination saga at Meta) highly optimized, and frequently changed. No human decision-maker at Meta chose to use… pictures of schoolgirls to advertise one of its social media platforms to a 37-year-old man—the ML system learned to do it.

Rob uses a win/win example (a car shopper buying a better car) but this is the same ML that will match precious metals scams to senior citizens, gambling ads to likely problem gamblers, and different campaign promises for the same politician to people with different positions on issues. (In some ways, the design of Meta’s ad system benefits deceptive advertisers over legit ones.) in the future, deceptive PACs won’t have to do false-flag ads for the other side manually. If they work, Meta’s systems will create and run them automatically.

According to the letter of the law here in California, Meta and Google are required to disclose their inferences to the person affected—but their lawyers are many and the system has been designed not to create obvious data points that would come out quickly in a future case. A regulatory agency that’s designed for obvious tracking that you can see in the browser’s dev tools is not going to be prepared to bring this stuff into compliance and will need to level up. And regulators can’t do it alone. Regulators, journalists and researchers, and private sector lawyers all have unique capabilities for addressing the surveillance problem, and regulators need to re-focus on areas where they can be a force multiplier for the other two.

IMHO a useful starting point would be to start attaching disclosure requirements to licensing processess that are already in place. For example, gambling apps should be required to disclose the targeting inferences used for their ads—and the big platform companies would need to provide adequate disclosures in order to be able to keep those apps as advertisers. More: Do you have a license for that robot bookie?