The Recommendation Algorithm That Runs Hollywood: How AI Is Deciding What Gets Made and What Gets Buried

By Stuart Kerr, Technology Correspondent, LiveAIWire

The AI recommendation algorithm behind Netflix now decides more of what you watch than any editor, critic or programming executive ever did. Netflix’s own long-standing figure, first disclosed by product executive Todd Yellin and documented in New America’s case study of the platform, holds that roughly 80 percent of what members watch is driven by algorithmic recommendation rather than active search or editorial curation. That single statistic represents a shift in cultural authority, from commissioning executives and artistic directors who chose what to elevate based on judgements about quality, to an AI recommendation algorithm trained on aggregate engagement data that surfaces what historical patterns predict will hold attention.

The same system that tells a subscriber what to watch next is now shaping what studios commission, what networks greenlight and what streaming platforms renew. AI has not replaced Hollywood’s decision-makers. It has restructured the information those decision-makers see and the success criteria they are held to, in ways that are quietly changing the range of stories being told. Understanding how an AI recommendation algorithm actually makes its decisions, and what it systematically rewards or buries, is essential context for anyone trying to understand why the entertainment landscape of 2026 looks the way it does.

How the AI Recommendation Algorithm Actually Decides

Streaming recommendation systems are machine learning models trained primarily on watch data, what content is played, for how long, whether viewers finish it or abandon it, and what they watch next. These models learn correlations between content characteristics and engagement outcomes, then surface content predicted to maximise engagement for each individual user. A viewer who has watched several crime thrillers is shown more crime thrillers. The content an AI recommendation algorithm surfaces most readily shares identifiable traits: familiar genre patterns, high-recognition talent, and a narrative shape that historically correlates with completion rather than abandonment.

The content that gets buried shares the opposite traits. Slow-burn storytelling that requires extended investment before payoff, unfamiliar cultural contexts that ask something of the viewer, debut or low-recognition creative talent, and stylistic ambition that departs from established patterns all face a structural disadvantage inside a system optimised for predicted completion, whatever their eventual critical or cultural value.

From Recommendation to Commissioning

The influence of recommendation data on commissioning decisions has been a documented feature of streaming-era production since at least 2020. What has changed is the sophistication of the predictive models studios apply at the greenlight stage, before a single frame is shot. Systems now analyse scripts, cast combinations, genre positioning and comparable title performance to generate predicted engagement scores that feed into development decisions. Netflix, Amazon and Apple have all been reported to use predictive performance modelling as part of that process, though the specific weight given to model output versus other commissioning criteria is not publicly disclosed by any of the three.

The consequence is a feedback loop. The AI recommendation algorithm identifies which content performs well on engagement metrics. That data informs predictive commissioning models, which favour content resembling what has already performed. The commissioned content performs well on the metrics it was designed to match, and the cycle reinforces itself.

The range of content that gets made narrows toward the centre of the distribution, toward what has worked before, measured by engagement metrics that may not capture the full value that risk-taking storytelling creates. No single executive decides to narrow the catalogue this way. The narrowing is an emergent property of an AI recommendation algorithm and a commissioning process feeding each other the same signal, repeatedly, without anyone in the loop necessarily intending the aggregate effect.

What Gets Buried, and What the Evidence Actually Says

The creative consequences of algorithmic commissioning are most visible in what disappears. A slow-burn prestige drama that generates intense loyalty among a smaller audience but a lower completion rate than a procedural crime show faces a structural disadvantage even if its critical recognition and cultural impact exceed what its engagement metrics suggest. An international-language film from an unfamiliar market faces the same disadvantage, because viewers conditioned to content in their own language are statistically less likely to complete it in a first session, regardless of quality.

The academic evidence on whether streaming genuinely narrows what people watch is more contested than the popular filter bubble narrative suggests, and an honest account of the AI recommendation algorithm’s effect on culture needs to hold both findings at once. A 2025 study by Samuel Coavoux and Abel Aussant, published in Sociological Science, found that streaming platform use in France was associated with a measurable increase in the diversity of content people consumed across music, film and television, not a decrease. The same study found that this diversity gain was not evenly distributed. It reinforced existing social gaps, with the benefit concentrated among viewers who already had the cultural capital to seek variety out.

That finding complicates the simple story that algorithms flatten taste. What it does not resolve is the commissioning-side effect this article is primarily about. Even if recommendation systems can surface a wider catalogue to viewers already inclined to explore it, the same systems can still discourage a studio from ever making the slow-burn, unfamiliar or debut-driven work in the first place, because the predicted engagement score at the greenlight stage is calculated before that exploratory viewing behaviour ever has a chance to occur.

The Creator Response to the AI Recommendation Algorithm

Writers, directors and performers working in streaming have taken different paths through the world the AI recommendation algorithm has created. Some have embraced the data, using audience analytics to understand which choices engage viewers. Others have pushed back directly against the idea that creative work worth making should be constrained by what a model predicts against historical baselines.

Our own analysis of why AI-generated movies fail to replace human storytelling found that AI systems trained on existing cinema can reproduce its surface patterns without the specific human perspective that makes the best films worth the hours they demand. That same limitation applies to a recommendation model predicting what audiences will finish just as much as it applies to a model trying to generate a script from scratch.

The 2023 writers’ and performers’ strikes, which produced explicit contractual limits on AI use in script development and new consent requirements around digital replicas, were partly a response to this same dynamic. The concern was not only generative AI writing scripts. It was the way AI-powered performance analytics were being used to justify commissioning and compensation decisions that writers whose shows drew strong critical response but weaker engagement scores than comparable titles had no way to contest.

The Cultural Stakes of an AI Recommendation Algorithm Choosing What Gets Made

The stakes here extend beyond any single show or film. Entertainment is one of the mechanisms a culture uses to develop empathy across difference and to surface experiences that would otherwise stay invisible to the majority. If an AI recommendation algorithm systematically surfaces familiar experiences to familiar audiences and starves unfamiliar ones of the engagement data a commissioning model needs to take a chance on them, that function is compromised in ways aggregate metrics will not reveal on their own.

The counterargument, that viewers are simply showing what they want through engagement data, has merit but is incomplete. Preferences are shaped by exposure. A viewer who has never been shown Korean social drama by the recommendation system has not expressed a preference against it. They have not encountered it. The algorithm’s role in shaping exposure means engagement data reflects the algorithm’s own prior choices as much as it reflects genuine viewer preference, a circularity that undercuts the claim that algorithmic curation is simply serving demand that already existed.

The International Market Distortion

The effect on non-English language content shows the dynamic most clearly. Netflix’s global catalogue spans dozens of languages, but the recommendation system surfaces that content based on predicted engagement, which is itself a function of historical viewing patterns. Subscribers who have not previously watched Korean drama or French auteur cinema have no historical pattern for the model to predict against, so those categories are rarely prominently recommended to them, regardless of how much they might enjoy the content once exposed to it.

Squid Game is the exception that proves the rule. The Korean-language thriller broke through to become one of Netflix’s most-watched series ever, but only through a combination of algorithmic promotion in specific markets, social media word of mouth that bypassed the recommendation system entirely, and marketing investment that overrode the platform’s normal prediction-based promotion logic. The lesson the industry drew was not that international content is systematically underexposed by the recommendation model. It was that international content can succeed given enough promotional support, which most international productions never receive.

The Institutional Response

SAG-AFTRA and the Writers Guild of America used the 2023 strikes to secure the contractual AI protections summarised in the WGA’s own 2023 agreement, which limit AI use in script development and require consent before a writer’s work can be used to train a generative model. Those protections address generative AI directly. They do not yet cover the separate question of how an AI recommendation algorithm’s engagement data shapes which scripts get greenlit in the first place, which is the gap unions are now examining.

Our reporting on Bryan Cranston’s fight with OpenAI over Sora 2 consent controls found that organised, specific professional demands can move a platform’s governance faster than legislation, a template unions are now studying for that broader commissioning question. A parallel case study is the landmark ruling that forced an AI platform to remove dozens of unauthorised voice clones built from French dubbing performers, which shows the same mechanism, a specific, technically coherent demand rather than a general objection to AI, succeeding where broader appeals to fairness had stalled.

The core demand in each case is the same: creative workers whose output is evaluated primarily by completion rates and engagement scores rather than by critical response or cultural significance need to understand the specific metrics an AI recommendation algorithm is applying to their work, so they can negotiate around them rather than guess at them.

The Monoculture Risk

The deepest concern about an AI recommendation algorithm shaping cultural production is not about any individual show or genre. It is about what happens when the feedback signal shaping what gets made becomes dominated by engagement metrics collected from recommendation-mediated viewing. A parallel dynamic is already documented in text: our coverage of the model collapse risk facing AI-generated content found that AI systems trained increasingly on smoothed, AI-shaped output progressively lose the diversity, edge cases and novel turns of phrase present in the human-generated material they were trained on. A commissioning system trained increasingly on the output of its own past recommendations risks the same compounding effect, narrowing toward an average that looks safer with every cycle.

Breaking Bad was renewed by AMC despite low initial ratings because executives believed in the creative vision. Game of Thrones was greenlit as an expensive fantasy adaptation with no comparable precedent to model against. Neither decision could have been made by a system optimising current engagement predictions, because the engagement data for work that redefines what audiences want does not exist until after the work has already been made, distributed and allowed to find its audience on its own terms. An AI recommendation algorithm can only ever predict from what has already happened. The shows that change what audiences want next, by definition, have not happened yet.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.