Beyond the ModelAI · systems · incentives · consequences

Editions / The Invisible Programme Controller

Edition 2 · 04 / 08

The Algorithm Has Become the Programme Controller

Recommenders perform an editorial function at immense scale, but their choices are governed differently from a broadcast schedule.

By Dr Gary Marrs PhD · · 3 min read

Recommender systemsMediaAccountability
A recommendation engine turning one media catalogue into many personalised sequences
Article summary

A programme controller decides what appears next. The role connects a broadcaster’s purpose to a schedule: this programme, for this audience, at this time, followed by that one.

Recommendation systems now perform a comparable function. They select from a large supply of eligible material, rank the alternatives and assemble a sequence. The difference is that they do it separately for each person, continuously and at enormous scale.

Prediction becomes an editorial act

A recommender may be described as a prediction system. It estimates what someone is likely to click, watch, enjoy or find useful. But once the prediction determines prominence, it becomes an intervention in the person’s information environment.

The user can only respond to what receives exposure. Material that is never shown has no chance to be selected. The resulting behaviour is then used as evidence for later predictions. Recommendation is therefore neither a neutral mirror of preference nor a conventional editorial decision. It is a feedback process in which selection and observation affect each other.

The 2016 paper describing YouTube’s learned recommendation architecture explains a two stage process of candidate generation and ranking at that time. It does not reveal the objectives or safeguards of the current system, but it illustrates the industrial scale at which ranking turns a vast catalogue into an individual sequence.

Mandates make purpose visible

The BBC’s public purposes include providing impartial news, supporting learning, showing creative and distinctive work, and reflecting the United Kingdom and the wider world. Sesame Street was built around an explicit educational purpose. In both examples, attention is valuable because it serves an independently stated aim.

A commercial recommender also has several objectives and constraints. Safety, satisfaction, relevance, quality and business goals may all contribute. The problem is not that a system uses engagement. It is that the public often cannot tell which purposes govern its choices, how conflicts are resolved or whether the selected measures represent the user’s interests.

One controller, millions of schedules

The analogy has limits. A broadcaster selects a shared schedule. A recommender reacts to individual behaviour and available content, and its output may be difficult even for the platform to summarise as a single editorial line.

That makes accountability more important. A traditional schedule can be inspected. A personalised exposure trajectory disappears unless it is recorded and made available for evaluation.

We should ask services to explain:

  • the important objectives that influence ranking;
  • the signals used as proxies for user interests;
  • the constraints applied to sequences as well as individual items;
  • how material system changes are reviewed;
  • and how users and independent researchers can examine cumulative exposure.

The algorithm has not replaced editorial judgement with an absence of judgement. It has embedded judgements in objectives, data, eligibility rules and ranking logic. Those choices deserve the scrutiny we once applied to the programme controller.

References