Hello everyone!
I’m Miko Pawlikowski, and in this #56 episode of The Hockey Stick Show, I sat down with Luca Belli, AI safety lead at Spring Health and author of the upcoming book Hidden Influences.
Luca spent more than five years at Twitter, where he co-founded the Machine Learning Ethics, Transparency, and Accountability team. Our conversation explored something most of us interact with every day without thinking much about it: recommendation algorithms.
They decide what appears in our feeds, what products we discover, which people we meet, and increasingly, how we understand what is happening in the world.
How Much of Your Life Is Chosen by an Algorithm?
Probably more than you think.
Social media has become one of our main gateways to the internet. We use it to keep in touch with people, read the news, discover ideas, and make sense of what is happening around us.
Almost everything we see there has been selected by a recommendation system.
But the influence goes much further than social media.
Dating apps recommend people we might want to meet. Event platforms suggest where we should go. Shopping sites decide which products are worth showing us. Those recommendations can eventually affect who we meet, what we buy, what we become interested in, and what we do offline.
The technology has become so deeply embedded in everyday life that we often stop noticing it.
Generative AI Is Only Part of the Story
Much of the current debate about AI focuses on generative models.
Fake images. AI-generated video. Misinformation.
Luca pointed out that creating misleading content is only part of the problem.
For that content to matter, somebody has to see it.
And very often, the mechanism deciding who sees it is a recommender system.
A fake video sitting somewhere on the internet has limited impact. A system that decides it is engaging and pushes it in front of millions of people is a different problem entirely.
Recommendation algorithms have been influencing information distribution for years, long before generative AI became the centre of the conversation.
The Algorithm Is Learning From You. You’re Learning From It.
One of the most interesting parts of our conversation was the difficulty of separating our own preferences from the systems around us.
Recommendation systems learn continuously from our behaviour.
You click something. You watch a video. You follow someone. You like a post.
Those actions become signals that influence what you see next.
But the influence also runs in the opposite direction.
An algorithm might repeatedly expose you to a topic you had never previously considered. Eventually you become interested in it.
Was that interest always there, waiting to be discovered?
Or did the recommendation system help create it?
There isn’t a clean boundary.
Luca described recommendation systems as socio-technical systems. You cannot properly study the algorithm without also studying the people interacting with it.
The user changes the system, and the system changes the user.
That feedback loop is part of what makes recommendation algorithms so powerful.
There Is No Correct Recommendation
Building a system that recognises whether an image contains a cat or a dog has something recommendation systems don’t: a ground truth.
Most people can agree whether the image contains a cat.
But what is the objectively correct recommendation for a person?
There isn’t one.
Something I find interesting might be completely irrelevant to you. Even my own preferences change depending on the situation.
The things I want to read in the morning might be completely different from what I want after a long day at work.
That creates a fundamental problem.
If there is no universal definition of a good recommendation, what exactly should these systems optimise for?
Clicks?
Time spent?
Satisfaction?
Something else entirely?
There is no simple gold standard against which a recommendation algorithm can be measured.
And that makes questions about whether one is behaving “correctly” much harder than they initially appear.
Hidden Influences
Luca’s book Hidden Influences grew out of his work studying responsible AI and the societal impact of recommendation systems.
At Twitter, he worked particularly on algorithmic amplification and the amplification of political content.
After Twitter was acquired in 2022, he decided it was time to move on. The book became a way to put several years of thinking and research into one place and, as he put it, get some closure before starting the next chapter.
It isn’t a manual for engineers trying to build a recommender system from scratch.
Instead, Luca wants to explain what these systems do, how they interact with people, and why their social consequences can be difficult to predict.
One audience he would particularly like to reach is policymakers.
Regulating recommendation algorithms requires understanding what actually happens under the hood, rather than treating “the algorithm” as a mysterious black box.
Moderation Gets Complicated Fast
Recommendation and moderation are closely connected.
If a platform is deciding what people should see, it also has to decide what they should not see.
That sounds straightforward until you try to write rules that work globally.
Even something seemingly simple like banning nudity requires a precise definition of what counts as nudity. Cultural expectations differ between countries, communities, and contexts.
Then there is the problem of who is doing the moderation.
A moderator may be reviewing content created in a culture they know very little about, intended for an audience with an entirely different set of assumptions.
Luca described this as a form of audience collapse.
Something originally created for one audience suddenly reaches another audience without the context required to interpret it properly.
That is one reason global moderation policies become extraordinarily difficult once you get into the details.
Taking Back Some Control
Towards the end of our conversation, I asked Luca for something practical.
If recommendation systems are designed to keep our attention, what can somebody actually do about it?
His answer was refreshingly simple.
Use limits.
Most phones now allow you to set daily time limits for individual apps. Luca uses them himself because they interrupt the moment when a few minutes of scrolling quietly turns into an hour.
The other recommendation was to become more deliberate about what you consume.
There is nothing wrong with watching something funny or completely pointless.
The question is whether you intended to spend three minutes doing it or somehow ended up spending three hours.
Simply recognising that many of these systems are optimised for engagement changes the way you interact with them.
Final Thoughts
Recommendation systems are easy to overlook precisely because they work so smoothly.
They sit between us and huge amounts of information, filtering the world into something manageable.
That can be incredibly useful.
But those same systems influence what we discover, what we pay attention to, who we interact with, and sometimes even what we believe.
Hidden Influences attempts to remove some of the mystery around how that process works and give people better tools for thinking about their relationship with algorithms.
The book is already available through Manning’s Early Access Program, where readers can access completed chapters while Luca continues writing it.
You can follow Luca and his work at lucab.phd or find him on LinkedIn.
It was a pleasure having Luca on the show, and I hope you enjoy the episode as much as I enjoyed the conversation.
Thanks for reading, and see you in the next one.
