Breaking out of the bubble

• On the boundless catalogues of streaming platforms, we need help to find something relevant. The problem arises when probability becomes the only criterion and personalisation hands us back an ever narrower version of our own taste.

All my life I have steered clear of horror films. I don’t like watching films that I know will keep me up at night. So when The Substance, the film starring Demi Moore, came out, my first instinct was to give it a miss. I simply wasn’t going to watch it. But my colleagues at the office insisted. Knowing me, they sensed there was something in that story I might like.

I trusted them and, without watching the trailer or doing much research, I pressed play and what followed was my best cinema experience of 2025. Something that would never have happened if I had relied on what an algorithm offered me, based on the things I usually do like to watch.

What a recommendation tries to predict

When a platform recommends a film to us, it usually combines multiple signals: what we have watched, searched for or abandoned; what users with similar behaviour choose; how popular a title is; or the features and tags it shares with others. Not all systems work the same way but, put simply, they pursue the same goal: reducing uncertainty and anticipating what we are most likely to choose.

That precision is useful. In a boundless catalogue we need help to find something relevant. The problem arises when probability becomes the only criterion: if everything we are shown is inferred from what we have already done, personalisation risks handing us back an ever narrower version of our own taste.

Eli Pariser called this mechanism the “filter bubble”, a term that became popular above all in relation to social media, but which describes just as precisely what can happen to any recommendation system: the better it learns from us, the more it confirms who we already are, and the less room it leaves for something different to reach us.

Hyper-personalisation, now driven by artificial intelligence, can take that logic much further. Systems will know not only what we have watched, but how long we hesitate before choosing it, at what point we abandon it or which elements caught our attention. The more precise that learning becomes, the more comfortable the experience will be. And, at the same time, the more quietly the world on offer to us may shrink.

Towards an algotorial model

On platforms with a strong editorial identity, editorial work starts long before a film appears on a screen. It lies in deciding which titles to acquire and why; in choosing the right moment to release them; in writing a synopsis that helps people understand what makes them special; in connecting them with other works, themes and filmmakers; and in deciding in which collection, row or spot on the homepage they can best find their audience.

It is ongoing, handcrafted work. The team watches the films, interprets what is happening in the world, listens to the community and builds connections that are not always obvious.

When a row tells you “Based on what you’ve watched”, the algorithm uses your history to identify a starting point. But the films it shows you are not necessarily there because other users followed the same path. They are there because someone on the editorial team has manually linked them to what you watched. This is called algotorial: the symbiosis between the algorithm and the work of the editorial team, in which neither replaces the other and the recommendation would not make the same sense without both.

Someone has decided that two films are in dialogue with each other through their theme, their tone, their cinematic language, a shared emotion or the historical moment they speak from. Likewise, our collections are not automatically generated groupings: they are journeys built by people. The algorithm can decide which one to show you and when; it does not invent the logic within it.

Michael Bhaskar put it clearly in Curation: the great challenge of our time is no longer producing content, but selecting what is worthwhile amid an ever-growing excess. According to Gracenote’s 2025 State of Playreport, more than 45% of streaming viewers find the glut of services and content overwhelming. And it is not just a feeling: viewers spend an average of 14 minutes searching for something to watch, a figure that rises above 16 minutes among 18 to 34-year-olds.

That selection needs judgement, and technology is what allows it to reach each person on a scale that would otherwise be impossible.

A map built by people

It is worth pausing on what that symbiosis means exactly. It is not about pitting an editor against an algorithm, nor about splitting decisions equally between them. It is a two-layer architecture: editorial judgement builds the relationships and the algorithm personalises access to them.

People create the map: they select, contextualise and connect the works. Technology uses what it knows about each user to decide at which point on that map they might start exploring it. The difference may seem small, but it changes the nature of the recommendation.

A model based mainly on aggregate behaviour says: “People who watched this film also watched this one.” The algotorial model can say: “Since you watched this film, we’re showing you another one that our team has linked to it for a specific reason.” The algorithm knows where you are starting from; editorial judgement suggests how far you could go.

That is why the algotorial approach is not the opposite of personalisation or hyper-personalisation. It can use both. What changes is who has built the universe of relationships being personalised, and for what purpose: it is not just about correctly guessing what you would probably choose, but about combining relevance, context, diversity and discovery.

Editorial judgement also gets it wrong sometimes: it makes poor choices, insists on a connection that doesn’t quite work, backs a title that fails to find its audience. The difference is not infallibility, but that when it fails, there is someone who can review it, correct it or explain why it was done, something an opaque system does not always offer. And that, ultimately, is why the algotorial approach can generate something that a pure algorithm cannot achieve on its own: trust. We don’t trust a system that only predicts; we trust a judgement that has a signature, that can get things wrong and answer for it. That signature is what, further down the line, gives us permission to take risks.

Beyond what you expect

What happens when each of us sees our universe shrink through the cumulative effect of an algorithm, when what we choose to watch, listen to or enjoy always reflects the worldview we already had, our certainties and convictions? It is not only individual taste that is impoverished: we can also become less open to difference, on and off the screen.

That is exactly what the signature behind the algotorial approach tries to counterbalance: without someone to answer for the recommendation, there is nothing to temper that logic of always reducing uncertainty.

The alternative is not to give up on data; it is to use data to find our way through abundance without handing it the whole decision. The algorithm brings scale and relevance; editorial judgement brings context, unexpected connections and the possibility of going beyond what we already know we like.

According to Deloitte’s 2025 Digital Media Trends survey in the United States, 53% of Gen Z and millennial respondents say they get better film and series recommendations through social media than through the streaming platforms themselves. In the same vein, the Gracenote report adds that more than a quarter of streaming consumers find those recommendations unhelpful for finding content.

I read that figure like this: young people have not stopped trusting human recommendation. They have found a place where that recommendation still has a face, a name and a conversation behind it, and that place is not always the streaming platform.

That, I believe, is where the real opportunity lies: not in competing with the immediacy of a viral video, but in building a community that curates, comments and dares to recommend what isn’t obvious. It is the same human signature we were talking about, but carried over to the voice of our audience too.

When you trust someone, whether an editorial team, a community or a friend whose taste you respect, you dare to go somewhere you would never have reached on your own. A system trained to reduce uncertainty will tend to steer you towards a safe choice. Trust can do the opposite: give you permission to take risks.

I would never have found my way to The Substance on my own. It took specific people to recognise a connection that my viewing history didn’t contain. That is what the algotorial approach should make possible at scale, and what we work on every day: starting from what we already know about you to open a door to something you don’t yet know might interest you.