The streaming algorithm destroyed cinema culture.
You've read the version. The Netflix recommendation engine optimises for engagement, engagement correlates with familiar comfort, familiar comfort drives away the challenging and the unfamiliar, the challenging and the unfamiliar are where the best films live, therefore the algorithm is killing the best films.
This is, with one significant revision, correct. The revision is: the algorithm doesn't create the problem. It exploits a problem that exists independently of any algorithm, which is the problem of passive consumption.
WHAT THE ALGORITHM ACTUALLY DOES
The Netflix algorithm is optimised for a specific metric: engagement, measured in minutes watched. A film you watch for ninety minutes and find deeply uncomfortable registers identically in the algorithm to a film you watch for ninety minutes and find deeply pleasurable. What the algorithm can't measure is what you got from the experience - whether the discomfort was productive or whether the pleasure was numbing.
The algorithm surfaces the film most likely to produce the next ninety minutes of engagement, not the film most likely to produce the most valuable ninety minutes of engagement. Those are different questions with different answers, and the algorithm only knows how to answer the first one.
THE PASSIVITY PROBLEM
The passivity problem is this: most people open a streaming platform without knowing what they want to watch. They open the interface and let the interface decide. The algorithm, optimising for engagement, serves them the content most likely to get them to click and keep watching - which is content that resembles things they've already watched, that has high production values and familiar genre mechanics, and that is immediately accessible rather than requiring patience or unfamiliarity.
This is not the algorithm's fault in any meaningful sense. The algorithm is doing what it's designed to do. The viewer opening the platform without a plan has opted into a system that will make decisions for them in the direction of the system's interests rather than the viewer's interests. The system's interest is engagement. The viewer's interest, presumably, is watching something worth watching. Those interests don't reliably align.
THE INTENTIONAL VIEWING PRACTICE
The antidote to the algorithm is not a better algorithm. It is the intentional viewing practice - knowing what you want to watch before you open the platform, using the platform as infrastructure for the plan rather than as the source of the plan.
The specific practice: before opening any streaming platform, spend sixty seconds on Letterboxd. Not scrolling - going directly to a watchlist you've built in advance. Choose the film. Open the platform with the film already decided. The algorithm doesn't get to participate.
This sounds trivial. The engagement data suggests it isn't. Netflix's own internal research has found that users who engage with the platform's discovery features watch more films per week than users who browse aimlessly. The directional watching habit produces more watching, not less - because the friction of the passive discovery process (scrolling through options without deciding) is itself the thing that leads to 'nothing to watch' syndrome and no watching at all.
THE MUBI MODEL AS AN ALTERNATIVE
MUBI's programming model is the algorithmic alternative fully realised: editorial curation by humans with a perspective, building a thirty-film catalogue chosen for quality and formal significance rather than for engagement optimisation. The algorithm doesn't get to decide what's in the MUBI catalogue. People with aesthetic commitments do.
The limitation of the MUBI model is that editorial curation by humans with a perspective reflects the perspective of the humans doing the curation - which produces its own blind spots and its own defaults. The virtue of the model is that the blind spots are human blind spots, driven by taste and aesthetic commitment rather than by engagement metrics. The distinction matters. Taste can be argued with. An engagement metric can't.
The antidote to the algorithm is not a better algorithm.