Why we will never have a recommendation algorithm

In the summer of 2021, a member named Liesbeth sent us an email. She'd been with PölderPlay since almost the beginning — she was in the second intake, March 2019 — and her question was simple: "Have you ever considered adding a 'because you watched X, try Y' feature?"

We wrote back a longer reply than she probably expected. We explained that no, we hadn't considered it, and that we were fairly certain we never would. This essay is an expanded version of that reply.

What recommendation algorithms actually optimise for

Let's be precise about what we mean by "recommendation algorithm." We're talking about the collaborative-filtering and content-based systems that power the discovery interfaces on Netflix, Spotify, YouTube, and their successors. The ones that say: given what you've watched before, here are things we predict you'll also watch.

These systems are not optimising for your genuine enrichment. They are optimising for engagement — specifically, for the probability that you will click, watch, and continue watching. This is sometimes the same thing as enrichment. Often it is not.

The problem is structural: a recommendation system is trained on behavioural data (clicks, watch-time, completion rates, ratings). Behavioural data reflects what people did, which is shaped by what was available and surfaced, which was in turn shaped by previous recommendations. The system is trained on its own outputs. It converges on a local optimum that tends to be conservative and familiar — because familiar things get clicked, and clicked things get recommended.

This matters less for music, where a mood-matching system can be genuinely useful. It matters enormously for film, where the most important experiences are often with work you couldn't have anticipated wanting — work that challenges your assumptions about what film can do, that comes from a tradition you didn't know existed, that takes you somewhere uncomfortable before it takes you somewhere real.

How films are actually discovered

I've been watching films seriously for about twenty-five years. I keep a list. In the last decade, the films that have mattered most to me came from: a conversation at a festival; a footnote in an essay about something else; a recommendation from a director I interviewed; a programme note at a repertory cinema; a colleague mentioning something in passing in a way that made me curious.

None of them came from an algorithm. None of them would have.

This isn't because I'm unusual. It's because the films that matter most tend to be precisely the ones that fall outside the behavioural patterns that algorithms are trained to recognise. They are, by definition, unexpected. You cannot build a system to deliver the unexpected — you can only build the conditions under which it might happen.

What we do instead

We do four things that we think create genuine conditions for discovery:

Curator's notes. Every film in our library has an accompanying essay. Not a synopsis. Not a star rating and three genre tags. An actual piece of writing that situates the film, explains what's interesting about it formally or thematically, and says what you might find difficult or surprising. These are written by our curators — real people who have watched the film and thought about it.

A non-algorithmic homepage. Our browse page is not personalised. Everyone who opens PölderPlay sees the same front page, edited by our team. This is a constraint, but we think it's a productive one. It means the films we surface are films we have reasons to surface, not films that a model thinks you'll click.

Film Circles. Small discussion groups of 8–12 members watching one film over two weeks. The discovery that happens in Film Circles is word-of-mouth in its most effective form — a recommendation from someone who knows you, who knows why a particular film might matter to you specifically, who has reasons beyond "you both gave four stars to Amour."

The catalogue itself as argument. We think of our catalogue not as a database but as a statement about what's worth watching. The juxtapositions matter. A library that puts an obscure 1972 Dutch film alongside a recent Austrian debut is making an argument that they belong in the same conversation. That argument is itself a form of guidance.

The objection I find most interesting

The most thoughtful objection to our position is this: if you're serious about discovery, shouldn't you want to surface films that members haven't seen? And if so, isn't some form of filtering — even algorithmic filtering — necessary to do that at scale?

This is a real tension. With 1,240 films in the library, a new member can't read all the curator's notes before choosing what to watch first. Some form of filtering is happening whether we acknowledge it or not.

Our answer is that the filtering we do should be editorial, not behavioural. It should be based on curators' judgements about what matters and why, not on what previous members have clicked. And it should be transparent — when we put a film on the front page, you should be able to see why. When an algorithm surfaces something, you cannot.

We're aware this doesn't fully resolve the tension. A library of 1,240 films curated by eight people still contains the biases of those eight people. We try to be honest about this. Our European focus is a genuine focus, not a claim to universality.

A closing thought

Liesbeth, the member who prompted this essay, eventually wrote back to say she understood our position even if she didn't entirely agree with it. She's still a member. She's in a Film Circle discussing a 1967 Czech film that a co-curator recommended to her group.

She told me last month that it's the best thing she's watched in years. She never would have found it herself.

That's the whole argument.