Facebook doesn’t care about your data. Its AI does.

This week, I started making Facebook ads to promote my new course AI expert.

It was a while since I last played with FB’s ad manager, so I was welcomed with a bunch of new privacy-related updates. I ended up on an interesting journey to research the current status of web tracking, data privacy, and how AI glues all of this together. You’re about to read the gist of what I found out.

Before we start, let me quickly explain how Facebook uses data to show ads. Lots of people have a foggy idea of that process, so it’s a good idea to shed some light.

Put yourself in my shoes. I want to teach AI to people, and have a small audience on social media. If I want to reach more people, Facebook tells me that it has 2.8B humans available to read my message.

That’s a lot of humans. They’re actually too many, and I don’t want to reach out to everyone in the world, just to the people who are interested in learning AI. So here’s how Facebook helps me: it first asks me what is my objective (in my case, getting new students for my class). It then asks me to specify my audience, in terms of age, location and interests, and starts showing ads just to the people within that audience who have the highest likelihood of joining my class.

This last part is where AI comes into play (and why FB is so powerful). Even by specifying “people interested in AI”, I would end up spamming people who don’t care about learning a new technology. For instance, I may show my ad to people who just like sci-fi movies but don’t want to invest in learning AI. Both me, FB, and people want to avoid this from happening, so FB came up with a cool idea: they’ll let some AI alrgorithm to find the best people for my goal.

How do they do it? Well, modern AI is fuelled by data: Facebook needs to track who’s joining my class so it can match her with other people who are similar and likely to join as well. This is where cookies come into play. In this case, cookies do nothing more than tell Facebook “hey, you know that user who saw the ad two days ago? She joined the class!”. That’s it, neither Facebook nor I need to know anything more about people’s web activity. Facebook’s AI just need to know what’s working, so it can do more of it.

What’s fascinating about this is that Facebook’s AI has gotten so good that Facebook suggests to keep your targeting broad so its AI can do its thing undisturbed by our assumptions. In other words, their AI can target much better than a marketer (this reminds me of Computer Scientist Frederick Jelinek’s quote on the superiority of Machine Learning over hand-crafted rules: “Every time I fire a linguist, the performance of the speech recognizer goes up”).

Better targeting is arguably better for everyone. Without it I’d end up proposing my course to people who’re not interested in AI, probably annoying them while also driving my advertising costs up.Entire businesses are built around FB’s simple concept of using AI to find people similar to the ones who performed a certain action (joining a course, buying a pair of shoes, sign a political campaign). But the castle crumbles if Facebook’s AI can’t learn from who performed a certain action.

Well, the castle is crumbling.

The internet is very young, and in the past decades it’s been the Wild West. While some small businesses like mine are using it as a tool to find people interested in our message, some greedy companies found space in the lack of regulation to make money exploiting people’s data. This includes tracking without consent, tracking more than necessary, or selling users’ behaviour data to other companies.As people and their governments are starting to say “this is enough”, tech companies are making their steps as well. And some of these steps are pissing Facebook off.

The two biggest shifts in data privacy related to targeted advertising are Apple’s iOS 14 update and Google’s announcement that they’ll “stop targeting based on past user’s web browsing” (you’ll understand why I added the quotes later).

Apple’s iOS14 update will require apps who track people outside of that same app to ask for user permission. This is the example of students joining my course through a Facebook ad: they see the ad on Facebook, but they join it following a process on Safari. It’s also the example of virtually any e-commerce advertising on Facebook.

If a student denies Facebook the option of tracking her outside of its app, Facebook won’t know that she joined the course in Safari. And remember that this is a big issue for Facebook, because its AI can’t have positive examples to learn from (making its targeting worse, spamming people, driving advertising costs up).

This is so important for Facebook that they bought two full page ads on the New York Times explaining why this move will damage small businesses.

And notice how the whole problem here is not that Facebook cares about reading the actions you take outside of the app. The problem is that its AI needs to know whether the ads it served you worked (or not), so it can find people similar (or different) from you. This whole drama is not about getting your personal data to Zuck, but it’s about giving Facebook’s AI positive and negative examples to learn from. Facebook estimated that without these examples, advertising costs would go up by 60% as a result of poorer targeting.

Google is solving the same problem trough a different path. They renewed their commitment in the blogpost “ steps towards a more privacy-first web “, which will entail phasing out third-party cookies (like FB’s cookie on non-FB websites), and stop serving ads based on individual browsing behaviour.

Emphasis on “individual”. In this new “privacy first web”, Google won’t know what you do on the internet. Instead, **Google will deploy tiny AI models into each of our smartphones,**which will report to Google an aggregated view of our behaviour. Google will then use this view to match you to other people similar to you, without having to directly monitor your specific actions.

In other words, instead of bringing your data to a massive centralised AI model, Google wants to give tiny AI models to everyone’s smartphone so they can consume the data exactly where it’s generated. According to Google, this approach allows them to keep serving relevant ads, without your data ever leaving your phone and therefore preserving your privacy.

There are two things I find fascinating about this battle.

1. This is not about Facebook wanting your data to spy on you. This is about Facebook’s AI being hungry for data (in technical terms, “labels”) so it can learn what works and what doesn’t. Zuck doesn’t care about what you do, its AI model does.
2. Google’s elegant solution is fascinating, because it solves the problem of personal data-hungry AI with…more AI.

It’s tough to say how the web will look like in 10 years. One thing is for sure: AI is part of the problem, and an even bigger part of the solution.

Originally published at https://blog.gianlucamauro.com on March 12, 2021.