Streaming Redesigned Entertainment: Who Wrote the New Rules?

When I asked Gaurav Gandhi where the balance lay between human editorial judgment and algorithmic optimization, he reframed the question before answering it. The tension you are asking about, he said, does not exist in the way you have framed it. Human editorial judgment decides what gets made; algorithms decide what reaches whom. They are doing different things.

This a polished move, made by the most prepared interlocutor in this series. Gaurav arrived with data ready, examples ready, and reframings ready for the questions a careful student might ask. When a question gets reframed, what stays standing is the framing itself, and the framing carries values that deserve to be examined.

The system examined is the Indian streaming entertainment economy as Amazon Prime Video has helped shape it. Gaurav is Vice President, Asia Pacific and ANZ, at Prime Video, with almost twenty-five years of experience across streaming, broadcast, and media. He sees the transformation from inside the institutional actor that has helped author it. The question is what the transformation has built and what it leaves in the catalog without surfacing.

The default viewer the recommendation infrastructure was built around is the one whose taste the algorithm has learned to predict. They watch enough content for their preferences to become legible as data. They have the device, bandwidth, and subscription to stream regularly. They live in a market the platform has decided to serve at scale, with marketing budget behind it. They speak a language the recommendation model treats as a primary signal rather than as an edge case.

Gaurav says the person with the most power in the relationship is the viewer. The viewer with that power is the one whose preferences the algorithm has already learned. The viewer outside the model’s learning set, in a language the model treats as secondary, in a market the platform has decided not to optimize for, is not in the relationship Gaurav is describing. They are in the catalog, perhaps, but not in the recommendation layer that decides who reaches them.

Ιn this system the algorithm is not making a decision about who to exclude. It is making a decision about who to surface, and the surfacing decisions concentrate around the tastes the algorithm has already learned. Everyone excluded is excluded by the same mechanism: their content, their taste, or their creative work is not included in the training set the recommendation infrastructure was optimised on.

Mirzapur travelled to 180 countries. Panchayat travelled to 167. Twenty-five percent of the viewers for Indian originals are outside India. These numbers describe reach, and reach is the metric Gaurav’s industry rewards. What they also describe, without stating it, is which Indian originals travelled. Both featured shows are Hindi-language, urban, and expensively made. The Tamil, Telugu, Bhojpuri, and Marathi originals the platform also produces do not appear in the travel statistics because the recommendation infrastructure does not push them to non-regional audiences. An original that lives in the catalogue but never reaches a viewer beyond its language base has not been distributed. It has been archived. The word “reach” measures the surface. The catalogue underneath is a different data set.

Alongside the catalogue-level exclusion, there is an exclusion at the creator level. Gaurav says the technology has removed the wall around content creation, and he is right. A smartphone produces film-grade footage. YouTube and social platforms accept uploads at no charge. The barrier the technology removed was the distribution barrier. What replaced it, without anyone announcing the replacement, is the recommendation barrier. Anyone can upload. Almost no one gets surfaced. The new gatekeeper is less visible than the old one because no human is telling the creator that their work will not be broadcast.

The final exclusion concerns the person the system claims to be serving. Gaurav notes that the recommendation engine surfaces more of what the viewer likes and that this can produce a comfort zone. What the viewer likes was already shaped by what the algorithm surfaced first, and the surfacing shapes the next round of what they encounter, which shapes what they come to prefer, which trains the model further. Over enough time the recommendation engine is no longer serving a taste. It is authoring one and then reflecting it back. The viewer experiences autonomy. The algorithm experiences it as a closed loop.

What this system optimizes for is visible in the metrics it tracks publicly. Engagement, completion rates, retention, scale across markets, growth in originals development. These are the numbers Gaurav offered without prompting. They are the numbers the platform’s commissioning, marketing, and recommendation decisions are calibrated to increase.

What the system does not optimize for, or at least does not measure publicly, is linguistic and regional diversity in surfacing, viewer agency over the recommendation model, creator economic transparency, and the breadth of the catalog that actually reaches non-regional audiences. The 58% new-talent figure and the 70% women HODs figure are real and meaningful. They describe what gets made. They do not describe what gets surfaced.

Gaurav says what algorithms do: they shape toward us, not away from us. The frame treats the viewer’s interests as known, the algorithm as responsive to them, and the relationship as one of service. The frame does the same kind of work the word natural did in the previous episode. It dissolves the question of who built the model, whose tastes it was trained on, and which viewers, languages, and creators it treats as edge cases. The relationship is service-oriented in one direction only. The other direction is design.

Each redesign that follows begins with a number Gaurav offered, and asks what would change if that number had to be reported every year.

The first is language and regional diversity reporting in the recommendation stack, not just in the production slate. Platforms should publicly commit to and report on the linguistic and regional distribution of content in their recommendation outputs. The travel statistics for Mirzapur and Panchayat describe what the algorithm pushes and ignores what the algorithm archives. The anticipated response is that recommendation is a personalised function and cannot be aggregated without privacy compromises. It can. Editorial share is aggregated all the time. Surfacing share follows the same aggregation logic, minus the individual watch histories.

A second redesign concerns algorithmic taste transparency for viewers. Gaurav says the viewer has the most power. A redesign builds the interface that makes that power real: viewers should be able to see a simplified map of what the recommendation engine thinks they like and correct it through explicit preference declaration, not only through watch behaviour. The current model lets the viewer signal taste only by consuming. The redesign lets them signal it by stating it.

Annual diversity reports on the full production pipeline should also be created. The 58% new talent figure and the 70% women HODs figure are significant, and exactly the kind of statistic that benefits from systematic reporting. A redesign asks platforms to publish annual reports covering new versus established talent, language distribution, gender of key roles, regional origin, and the relationship between any of these and surfacing in the recommendation layer. Accountability through transparency rather than through good-faith citation.

Finally, a creator economics transparency standard. Streaming platforms should disclose, at minimum to creator representatives and unions, the data models used to determine licensing fees, renewal decisions, and global distribution allocations. Creators cannot negotiate fairly in a system where the valuation methodology is invisible to them. The standard does not require disclosing proprietary algorithms; it requires disclosing the inputs and weights that govern decisions creators are subject to.

Eight interviews in, I know the shape a difficult conversation takes when the person on the other side has thought carefully about the question I am about to ask. It arrives at me as a reframe. The tension I was reaching for does not exist in the form I named it. The framing I brought is corrected, gently, and a cleaner framing is offered in its place.

This is not a bad-faith move. It is what expertise sounds like when it is asked to explain itself to a nineteen-year-old with a microphone.

The reframe is where the audit’s work begins. Human editorial decides what gets made; algorithms decide what reaches whom. Both halves of the sentence are true. Placed side by side, they present the two functions as parallel rather than as entangled, and the entanglement is where the accountability sits. The commissioning meeting is looking at data the recommendation engine produced. The recommendation engine is trained on catalogues that commissioning meetings chose. Nothing in either room is a lie. What the sentence hides is that neither room is separately in charge of the outcome.

If there is a single observation the eight audits together let me make, it is this. The strongest guests were not the ones with the most defensive framings. They were the ones whose framings were most coherent. Coherence is what makes a system’s design values invisible to the people operating inside it, because coherent framings do not present themselves as choices. They present themselves as the way things are. The audit’s only job is to name the choices anyway.