A Public Interest Framework for Recommender Systems

Recommendation is now part of the infrastructure through which people encounter culture, news, education and politics. Governing it requires more than removing prohibited items. We also need to understand the sequences produced from material that remains eligible.

This framework offers eight questions for product teams, boards, regulators and researchers. It is a proposal for scrutiny rather than a claim that every service should produce the same answer.

1. Whose interests govern the system?

Identify the people affected by ranking and state whose interests the system is intended to serve. Separate user value, advertiser value, creator value, public value and the service’s commercial objectives. Do not hide conflicts inside a single measure of engagement.

Evidence to seek: approved objectives, responsible owners and records showing how conflicts are decided.

2. What does the system infer?

List the behavioural and contextual signals used to represent interest. Explain significant inferences and identify where prominence, prior exposure or limited alternatives may contaminate the apparent preference.

Evidence to seek: feature documentation, data provenance, retention rules and analysis of proxy validity.

3. What becomes eligible and prominent?

Content moderation determines what may appear. Ranking determines what is likely to be encountered. Review both layers and the routes through which borderline material can acquire reach without crossing a removal threshold.

Evidence to seek: eligibility policies, ranking constraints and distribution analysis across topics and groups.

4. What happens across the sequence?

Evaluate cumulative exposure, not only individual recommendations. Measure repetition, source diversity, intensity, session duration, pace, natural stopping points and the ability to move beyond an inferred interest.

Evidence to seek: exposure trajectory tests and monitoring for foreseeable escalation patterns.

5. Which outcomes count?

Engagement is observable but incomplete. Define the human outcome the service is trying to support and the limitations of each proxy. Use several measures where values conflict.

Evidence to seek: outcome definitions, validation studies, subgroup analysis and documented uncertainty.

6. What control does the person have?

People should be able to understand and alter important recommendation inputs. Useful controls include chronological alternatives, topic controls, history inspection, resets, stopping points and explanations that lead to an effective action.

Evidence to seek: usability research showing that different groups can find, understand and use the controls.

7. Who can test the claims?

Internal evaluation is necessary but insufficient where a system has broad social effects. Provide privacy preserving access that allows qualified independent researchers and regulators to test exposure, outcomes and unequal effects.

Evidence to seek: reproducible methods, access arrangements, publication of limitations and responses to contrary findings.

8. Who is accountable for change?

Recommendation systems evolve through models, objectives, data, interfaces and content supply. Define which changes require risk assessment, senior review and external notice. Retain enough evidence to reconstruct consequential decisions.

Evidence to seek: change logs, named decision makers, incident criteria, audit trails and meaningful public summaries.

A practical review

Apply the framework to one real user journey. Reconstruct the eligible set, the ranking decision, the resulting sequence and the feedback captured from it. Ask what the system learned, what the person could control and which evidence would reveal an unintended effect.

The aim is not to replace personalisation with a universal schedule. It is to make the new programme controllers explain their purpose, demonstrate their controls and remain accountable for the environments they create.